move source code to src/

13 files changed, 2943 insertions, 0 deletions

diff --git a/src/__init__.py b/src/__init__.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/src/__init__.py
diff --git a/src/benchmarks/bench_loop.c b/src/benchmarks/bench_loop.c
new file mode 100644
index 0000000..bc15808
--- /dev/null
+++ b/src/benchmarks/bench_loop.c
@@ -0,0 +1,87 @@
+#include <assert.h>
+#include <malloc.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+
+static size_t _rand() {
+	static __thread size_t seed = 123456789;
+	size_t a = 1103515245;
+	size_t c = 12345;
+	size_t m = 1 << 31;
+	seed = (a * seed + c) % m;
+		return seed;
+}
+
+typedef struct ThreadArgs {
+	double benchmark;
+	int allocations;
+	int max_size;
+} ThreadArgs;
+
+static void* malloc_then_write(size_t size) {
+	void* ptr = malloc(size);
+	// Write to ptr
+	/* *((char*)ptr) = '!'; */
+	return ptr;
+}
+
+static void read_then_free(void* ptr) {
+	// Read before free
+	/* char s __attribute__((unused)) = *((char*)ptr); */
+	free(ptr);
+}
+static void* test_thread_func(void* arg) {
+	ThreadArgs* args = (ThreadArgs*)arg;
+
+	for(int i = 0; i < args->allocations; i++) {
+		void* ptr = malloc_then_write((_rand() % args->max_size) + 1);
+		read_then_free(ptr);
+	}
+	return NULL;
+}
+
+int main(int argc, char* argv[]) {
+	pthread_t* threads;
+	int num_threads;
+	struct ThreadArgs thread_args;
+
+	if (argc < 4) {
+		fprintf(stderr, "Usage: %s <num threads> <num allocations> <max size>\n", argv[0]);
+		return 1;
+	}
+
+	num_threads = atoi(argv[1]);
+	thread_args.allocations = atoi(argv[2]);
+	thread_args.max_size = atoi(argv[3]);
+
+	threads = (pthread_t*)malloc(num_threads * sizeof(pthread_t));
+
+	for (int i = 0; i < num_threads; i++) {
+		if (0 != pthread_create(&threads[i], NULL, test_thread_func, &thread_args)) {
+			perror("pthread_create");
+			return 1;
+		}
+	}
+
+	for(int i = 0; i < num_threads; i++) {
+		if (0 != pthread_join(threads[i], NULL)) {
+			perror("pthread_join");
+			return 1;
+		}
+	}
+
+	if (argc == 5)
+	{
+		FILE* f = stdout;
+		if (strcmp(argv[4],"stdout") != 0)
+			f = fopen(argv[4], "w");
+		malloc_info(0, f);
+		if (strcmp(argv[4],"stdout") != 0)
+			fclose(f);
+	}
+
+	return 0;
+}
diff --git a/src/benchmarks/cache-scratch.cc b/src/benchmarks/cache-scratch.cc
new file mode 100644
index 0000000..2cb9b28
--- /dev/null
+++ b/src/benchmarks/cache-scratch.cc
@@ -0,0 +1,147 @@
+///-*-C++-*-//////////////////////////////////////////////////////////////////
+//
+// Hoard: A Fast, Scalable, and Memory-Efficient Allocator
+//        for Shared-Memory Multiprocessors
+// Contact author: Emery Berger, http://www.cs.umass.edu/~emery
+//
+// This library is free software; you can redistribute it and/or modify
+// it under the terms of the GNU Library General Public License as
+// published by the Free Software Foundation, http://www.fsf.org.
+//
+// This library is distributed in the hope that it will be useful, but
+// WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+// Library General Public License for more details.
+//
+//////////////////////////////////////////////////////////////////////////////
+
+/**
+ * @file cache-scratch.cpp
+ *
+ * cache-scratch is a benchmark that exercises a heap's cache locality.
+ * An allocator that allows multiple threads to re-use the same small
+ * object (possibly all in one cache-line) will scale poorly, while
+ * an allocator like Hoard will exhibit near-linear scaling.
+ *
+ * Try the following (on a P-processor machine):
+ *
+ *  cache-scratch 1 1000 1 1000000
+ *  cache-scratch P 1000 1 1000000
+ *
+ *  cache-scratch-hoard 1 1000 1 1000000
+ *  cache-scratch-hoard P 1000 1 1000000
+ *
+ *  The ideal is a P-fold speedup.
+*/
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "fred.h"
+#include "cpuinfo.h"
+#include "timer.h"
+
+// This class just holds arguments to each thread.
+class workerArg {
+public:
+
+  workerArg() {}
+
+  workerArg (char * obj, int objSize, int repetitions, int iterations)
+    : _object (obj),
+      _objSize (objSize),
+      _iterations (iterations),
+      _repetitions (repetitions)
+  {}
+
+  char * _object;
+  int _objSize;
+  int _iterations;
+  int _repetitions;
+};
+
+
+#if defined(_WIN32)
+extern "C" void worker (void * arg)
+#else
+extern "C" void * worker (void * arg)
+#endif
+{
+  // free the object we were given.
+  // Then, repeatedly do the following:
+  //   malloc a given-sized object,
+  //   repeatedly write on it,
+  //   then free it.
+  workerArg * w = (workerArg *) arg;
+  delete w->_object;
+  workerArg w1 = *w;
+  for (int i = 0; i < w1._iterations; i++) {
+    // Allocate the object.
+    char * obj = new char[w1._objSize];
+    // Write into it a bunch of times.
+    for (int j = 0; j < w1._repetitions; j++) {
+      for (int k = 0; k < w1._objSize; k++) {
+	obj[k] = (char) k;
+	volatile char ch = obj[k];
+	ch++;
+      }
+    }
+    // Free the object.
+    delete [] obj;
+  }
+
+#if !defined(_WIN32)
+  return NULL;
+#endif
+}
+
+
+int main (int argc, char * argv[])
+{
+  int nthreads;
+  int iterations;
+  int objSize;
+  int repetitions;
+
+  if (argc > 4) {
+    nthreads = atoi(argv[1]);
+    iterations = atoi(argv[2]);
+    objSize = atoi(argv[3]);
+    repetitions = atoi(argv[4]);
+  } else {
+    fprintf (stderr, "Usage: %s nthreads iterations objSize repetitions\n", argv[0]);
+    return 1;
+  }
+
+  HL::Fred * threads = new HL::Fred[nthreads];
+  HL::Fred::setConcurrency (HL::CPUInfo::getNumProcessors());
+
+  workerArg * w = new workerArg[nthreads];
+
+  int i;
+
+  // Allocate nthreads objects and distribute them among the threads.
+  char ** objs = new char * [nthreads];
+  for (i = 0; i < nthreads; i++) {
+    objs[i] = new char[objSize];
+  }
+
+  HL::Timer t;
+  t.start();
+
+  for (i = 0; i < nthreads; i++) {
+    w[i] = workerArg (objs[i], objSize, repetitions / nthreads, iterations);
+    threads[i].create (&worker, (void *) &w[i]);
+  }
+  for (i = 0; i < nthreads; i++) {
+    threads[i].join();
+  }
+  t.stop();
+
+  delete [] threads;
+  delete [] objs;
+  delete [] w;
+
+  printf ("Time elapsed = %f seconds.\n", (double) t);
+  return 0;
+}
diff --git a/src/benchmarks/cache-thrash.cc b/src/benchmarks/cache-thrash.cc
new file mode 100644
index 0000000..79242eb
--- /dev/null
+++ b/src/benchmarks/cache-thrash.cc
@@ -0,0 +1,134 @@
+///-*-C++-*-//////////////////////////////////////////////////////////////////
+//
+// Hoard: A Fast, Scalable, and Memory-Efficient Allocator
+//        for Shared-Memory Multiprocessors
+// Contact author: Emery Berger, http://www.cs.umass.edu/~emery
+//
+// Copyright (c) 1998-2003, The University of Texas at Austin.
+//
+// This library is free software; you can redistribute it and/or modify
+// it under the terms of the GNU Library General Public License as
+// published by the Free Software Foundation, http://www.fsf.org.
+//
+// This library is distributed in the hope that it will be useful, but
+// WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+// Library General Public License for more details.
+//
+//////////////////////////////////////////////////////////////////////////////
+
+/**
+ * @file  cache-thrash.cpp
+ * @brief cache-thrash is a benchmark that exercises a heap's cache-locality.
+ *
+ * Try the following (on a P-processor machine):
+ *
+ *  cache-thrash 1 1000 1 1000000
+ *  cache-thrash P 1000 1 1000000
+ *
+ *  cache-thrash-hoard 1 1000 1 1000000
+ *  cache-thrash-hoard P 1000 1 1000000
+ *
+ *  The ideal is a P-fold speedup.
+*/
+
+
+#include <iostream>
+#include <stdlib.h>
+
+using namespace std;
+
+#include "cpuinfo.h"
+#include "fred.h"
+#include "timer.h"
+
+// This class just holds arguments to each thread.
+class workerArg {
+public:
+	workerArg() {}
+	workerArg (int objSize, int repetitions, int iterations)
+	: _objSize (objSize),
+	 _iterations (iterations),
+	 _repetitions (repetitions)
+	{}
+
+	int _objSize;
+	int _iterations;
+	int _repetitions;
+};
+
+
+#if defined(_WIN32)
+extern "C" void worker (void * arg)
+#else
+extern "C" void * worker (void * arg)
+#endif
+{
+	// Repeatedly do the following:
+	//   malloc a given-sized object,
+	//   repeatedly write on it,
+	//   then free it.
+	workerArg * w = (workerArg *) arg;
+	workerArg w1 = *w;
+	for (int i = 0; i < w1._iterations; i++) {
+	// Allocate the object.
+		char * obj = new char[w1._objSize];
+		//    printf ("obj = %p\n", obj);
+		// Write into it a bunch of times.
+		for (int j = 0; j < w1._repetitions; j++) {
+			for (int k = 0; k < w1._objSize; k++) {
+				obj[k] = (char) k;
+				volatile char ch = obj[k];
+				ch++;
+			}
+		}
+		// Free the object.
+		delete [] obj;
+	}
+#if !defined(_WIN32)
+	return NULL;
+#endif
+}
+
+
+int main (int argc, char * argv[])
+{
+	int nthreads;
+	int iterations;
+	int objSize;
+	int repetitions;
+	
+	if (argc > 4) {
+		nthreads = atoi(argv[1]);
+		iterations = atoi(argv[2]);
+		objSize = atoi(argv[3]);
+		repetitions = atoi(argv[4]);
+	} else {
+	cerr << "Usage: " << argv[0] << " nthreads iterations objSize repetitions" << endl;
+	exit(1);
+	}
+
+	HL::Fred * threads = new HL::Fred[nthreads];
+	HL::Fred::setConcurrency (HL::CPUInfo::getNumProcessors());
+    
+	int i;
+  
+	HL::Timer t;
+	t.start();
+  
+	workerArg * w = new workerArg[nthreads];
+    
+	for (i = 0; i < nthreads; i++) {
+		w[i] = workerArg (objSize, repetitions / nthreads, iterations);
+		threads[i].create (&worker, (void *) &w[i]);
+	}
+	for (i = 0; i < nthreads; i++) {
+		threads[i].join();
+	}
+	t.stop();
+
+	delete [] threads;
+	delete [] w;
+
+	cout << "Time elapsed = " << (double) t << " seconds." << endl;
+}
diff --git a/src/benchmarks/cpuinfo.h b/src/benchmarks/cpuinfo.h
new file mode 100644
index 0000000..1ed1f36
--- /dev/null
+++ b/src/benchmarks/cpuinfo.h
@@ -0,0 +1,202 @@
+// -*- C++ -*-

+

+/*

+

+  Heap Layers: An Extensible Memory Allocation Infrastructure

+  

+  Copyright (C) 2000-2003 by Emery Berger

+  http://www.cs.umass.edu/~emery

+  emery@cs.umass.edu

+  

+  This program is free software; you can redistribute it and/or modify

+  it under the terms of the GNU General Public License as published by

+  the Free Software Foundation; either version 2 of the License, or

+  (at your option) any later version.

+  

+  This program is distributed in the hope that it will be useful,

+  but WITHOUT ANY WARRANTY; without even the implied warranty of

+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

+  GNU General Public License for more details.

+  

+  You should have received a copy of the GNU General Public License

+  along with this program; if not, write to the Free Software

+  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA

+

+*/

+

+

+

+#ifndef HL_CPUINFO_H

+#define HL_CPUINFO_H

+

+#if defined(_WIN32)

+#include <windows.h>

+#include <process.h>

+#else

+#include <unistd.h>

+#endif

+

+

+#if !defined(_WIN32)

+#include <pthread.h>

+#endif

+

+#if defined(__SVR4) // Solaris

+#include <sys/lwp.h>

+extern "C" unsigned int lwp_self(void);

+#include <thread.h>

+extern "C" int _thr_self(void);

+#endif

+

+#if defined(__linux)

+#include <sys/types.h>

+#include <sys/stat.h>

+#include <fcntl.h>

+#include <string.h>

+#include <unistd.h>

+#endif

+

+#if defined(__APPLE__)

+#include <sys/types.h>

+#include <sys/sysctl.h>

+#endif

+

+#if defined(__sgi)

+#include <sys/types.h>

+#include <sys/sysmp.h>

+#include <sys/sysinfo.h>

+#endif

+

+#if defined(hpux)

+#include <sys/mpctl.h>

+#endif

+

+#if defined(_WIN32)

+extern __declspec(thread) int localThreadId;

+#endif

+

+#if defined(__SVR4) && defined(MAP_ALIGN)

+extern volatile int anyThreadStackCreated;

+#endif

+

+namespace HL {

+

+/**

+ * @class CPUInfo

+ * @author Emery Berger <http://www.cs.umass.edu/~emery>

+ *

+ * @brief Architecture-independent wrapper to get number of CPUs. 

+ */

+

+class CPUInfo {

+public:

+  CPUInfo (void)

+  {}

+

+  inline static int getNumProcessors (void) {

+    static int _numProcessors = computeNumProcessors();

+    return _numProcessors;

+  }

+

+  static inline unsigned long getThreadId (void);

+  inline static int computeNumProcessors (void);

+

+};

+

+

+int CPUInfo::computeNumProcessors (void)

+{

+  static int np = 0;

+  if (!np) {

+#if defined(__linux) || defined(__APPLE__)

+    np = (int) sysconf(_SC_NPROCESSORS_ONLN);

+#elif defined(_WIN32)

+    SYSTEM_INFO infoReturn[1];

+    GetSystemInfo (infoReturn);

+    np = (int) (infoReturn->dwNumberOfProcessors);

+#elif defined(__sgi)

+    np = (int) sysmp(MP_NAPROCS);

+#elif defined(hpux)

+    np = mpctl(MPC_GETNUMSPUS, NULL, NULL); // or pthread_num_processors_np()?

+#elif defined(_SC_NPROCESSORS_ONLN)

+    np = (int) (sysconf(_SC_NPROCESSORS_ONLN));

+#else

+    np = 2;

+    // Unsupported platform.

+    // Pretend we have at least two processors. This approach avoids the risk of assuming

+    // we're on a uniprocessor, which might lead clever allocators to avoid using atomic

+    // operations for all locks.

+#endif

+    return np;

+  } else {

+    return np;

+  }

+}

+

+  // Note: when stacksize arg is NULL for pthread_attr_setstacksize [Solaris],

+// stack size is 1 MB for 32-bit arch, 2 MB for 64-bit arch.

+// pthread_attr_getstacksize

+// pthread_attr_setstackaddr

+// pthread_attr_getstackaddr

+// PTHREAD_STACK_SIZE is minimum.

+// or should we just assume we have __declspec(thread) or __thread?

+

+#if defined(USE_THREAD_KEYWORD)

+  extern __thread int localThreadId;

+#endif

+

+  // FIX ME FIXME

+  //#include <stdio.h>

+

+unsigned long CPUInfo::getThreadId (void) {

+#if defined(__SVR4)

+  size_t THREAD_STACK_SIZE;

+  if (sizeof(size_t) <= 4) {

+    THREAD_STACK_SIZE = 1048576;

+  } else {

+    // 64-bits.

+    THREAD_STACK_SIZE = 1048576 * 2;

+  }

+  if (0) { // !anyThreadStackCreated) {

+    // We know a priori that all stack variables

+    // are on different stacks. Since no one has created

+    // a special one, we are in control, and thus all stacks

+    // are 1 MB in size and on 1 MB boundaries.

+    // (Actually: 1 MB for 32-bits, 2 MB for 64-bits.)

+    char buf;

+    return (((size_t) &buf) & ~(THREAD_STACK_SIZE-1)) >> 20;

+  } else {

+    return (int) pthread_self();

+  }

+#elif defined(_WIN32)

+  // It looks like thread id's are always multiples of 4, so...

+  return GetCurrentThreadId() >> 2;

+#elif defined(__APPLE__)

+  // Consecutive thread id's in Mac OS are 4096 apart;

+  // dividing off the 4096 gives us an appropriate thread id.

+  int tid = (int) ((unsigned long) pthread_self()) >> 12;

+  return tid;

+#elif defined(__BEOS__)

+  return find_thread(0);

+#elif defined(USE_THREAD_KEYWORD)

+  return localThreadId;

+#elif defined(__linux) || defined(PTHREAD_KEYS_MAX)

+  // Consecutive thread id's in Linux are 1024 apart;

+  // dividing off the 1024 gives us an appropriate thread id.

+  return (unsigned long) pthread_self() >> 10;

+#elif defined(POSIX)

+  return (unsigned long) pthread_self();

+#elif USE_SPROC

+  // This hairiness has the same effect as calling getpid(),

+  // but it's MUCH faster since it avoids making a system call

+  // and just accesses the sproc-local data directly.

+  unsigned long pid = (unsigned long) PRDA->sys_prda.prda_sys.t_pid;

+  return pid;

+#else

+  return 0;

+#endif

+}

+

+}

+

+#endif

diff --git a/src/benchmarks/fred.h b/src/benchmarks/fred.h
new file mode 100644
index 0000000..b0198a7
--- /dev/null
+++ b/src/benchmarks/fred.h
@@ -0,0 +1,97 @@
+// -*- C++ -*-
+
+#ifndef HL_FRED_H
+#define HL_FRED_H
+
+/// A thread-wrapper of childlike simplicity :).
+
+#if defined(_WIN32)
+
+  #include <windows.h>
+  #include <process.h>
+
+#elif defined(__SVR4)
+
+  #include <thread.h>
+  #include <pthread.h>
+  #include <unistd.h>
+
+#else
+
+  #include <pthread.h>
+  #include <unistd.h>
+
+#endif
+
+typedef void * (*ThreadFunctionType) (void *);
+
+namespace HL {
+
+class Fred {
+public:
+
+  Fred() {
+#if !defined(_WIN32)
+    pthread_attr_init (&attr);
+    pthread_attr_setscope (&attr, PTHREAD_SCOPE_SYSTEM);
+#endif
+  }
+
+  ~Fred() {
+#if !defined(_WIN32)
+    pthread_attr_destroy (&attr);
+#endif
+  }
+
+  void create (ThreadFunctionType function, void * arg) {
+#if defined(_WIN32)
+    t = CreateThread (0, 0, (LPTHREAD_START_ROUTINE) *function, (LPVOID) arg, 0, 0);
+#else
+    pthread_create (&t, &attr, function, arg);
+#endif
+  }
+
+  void join (void) {
+#if defined(_WIN32)
+    WaitForSingleObject (t, INFINITE);
+#else
+    pthread_join (t, NULL);
+#endif
+  }
+
+  static void yield (void) {
+#if defined(_WIN32)
+    Sleep (0);
+#elif defined(__SVR4)
+    thr_yield();
+#else
+    sched_yield();
+#endif
+  }
+
+
+  static void setConcurrency (int n) {
+#if defined(_WIN32)
+#elif defined(__SVR4)
+    thr_setconcurrency (n);
+#else
+    pthread_setconcurrency (n);
+#endif
+  }
+
+
+private:
+#if defined(_WIN32)
+  typedef HANDLE FredType;
+#else
+  typedef pthread_t FredType;
+  pthread_attr_t attr;
+#endif
+
+  FredType t;
+};
+
+}
+
+
+#endif
diff --git a/src/benchmarks/larson.cc b/src/benchmarks/larson.cc
new file mode 100644
index 0000000..be8038f
--- /dev/null
+++ b/src/benchmarks/larson.cc
@@ -0,0 +1,744 @@
+#include <assert.h>
+#include <stdio.h>
+
+#if defined(_WIN32)
+#define __WIN32__
+#endif
+
+#ifdef __WIN32__
+#include  <windows.h>
+#include  <conio.h>
+#include  <process.h>
+
+#else
+#include <unistd.h>
+#include <sys/resource.h>
+#include <sys/time.h>
+
+#ifndef __SVR4
+//extern "C" int pthread_setconcurrency (int) throw();
+#include <pthread.h>
+#endif
+
+
+typedef void * LPVOID;
+typedef long long LONGLONG;
+typedef long DWORD;
+typedef long LONG;
+typedef unsigned long ULONG;
+typedef union _LARGE_INTEGER {
+  struct {
+    DWORD LowPart;
+    LONG  HighPart;
+  } foo;
+  LONGLONG QuadPart;    // In Visual C++, a typedef to _ _int64} LARGE_INTEGER;
+} LARGE_INTEGER;
+typedef long long _int64;
+#ifndef TRUE
+enum { TRUE = 1, FALSE = 0 };
+#endif
+#include <assert.h>
+#define _ASSERTE(x) assert(x)
+#define _inline inline
+void Sleep (long x) 
+{
+  //  printf ("sleeping for %ld seconds.\n", x/1000);
+  sleep(x/1000);
+}
+
+void QueryPerformanceCounter (long * x)
+{
+  struct timezone tz;
+  struct timeval tv;
+  gettimeofday (&tv, &tz);
+  *x = tv.tv_sec * 1000000L + tv.tv_usec;
+}
+
+void QueryPerformanceFrequency(long * x)
+{
+  *x = 1000000L;
+}
+
+
+#include  <stdio.h>
+#include  <stdlib.h>
+#include  <stddef.h>
+#include  <string.h>
+#include  <ctype.h>
+#include  <time.h>
+#include  <assert.h>
+
+#define _REENTRANT 1
+#include <pthread.h>
+#ifdef __sun
+#include <thread.h>
+#endif
+typedef void * VoidFunction (void *);
+void _beginthread (VoidFunction x, int, void * z)
+{
+  pthread_t pt;
+  pthread_attr_t pa;
+  pthread_attr_init (&pa);
+
+#if 1//defined(__SVR4)
+  pthread_attr_setscope (&pa, PTHREAD_SCOPE_SYSTEM); /* bound behavior */
+#endif
+
+  //  printf ("creating a thread.\n");
+  int v = pthread_create(&pt, &pa, x, z);
+  //  printf ("v = %d\n", v);
+}
+#endif
+
+
+#if 0
+static char buf[65536];
+
+#define malloc(v) &buf
+#define free(p) 
+#endif
+
+#undef CPP
+//#define CPP
+//#include "arch-specific.h"
+
+#if USE_ROCKALL
+//#include "FastHeap.hpp"
+//FAST_HEAP theFastHeap (1024 * 1024, true, true, true);
+
+typedef int SBIT32;
+
+#include "SmpHeap.hpp"
+SMP_HEAP theFastHeap (1024 * 1024, true, true, true);
+
+void * operator new( unsigned int cb )
+{
+  void *pRet = theFastHeap.New ((size_t)cb) ;
+  return pRet;
+}
+
+void operator delete(void *pUserData )
+{
+  theFastHeap.Delete (pUserData) ;
+}
+#endif
+
+#if 0
+extern "C" void * hdmalloc (size_t sz) ;
+extern "C" void hdfree (void * ptr) ;
+extern "C" void hdmalloc_stats (void) ;
+void * operator new( unsigned int cb )
+{
+  void *pRet = hdmalloc((size_t)cb) ;
+  return pRet;
+}
+
+void operator delete(void *pUserData )
+{
+  hdfree(pUserData) ;
+}
+#endif
+
+
+
+/* Test driver for memory allocators           */
+/* Author: Paul Larson, palarson@microsoft.com */
+#define MAX_THREADS     100
+#define MAX_BLOCKS  20000000
+
+int volatile  stopflag=FALSE ;       
+
+struct lran2_st {
+  long x, y, v[97];
+};
+
+int     TotalAllocs=0 ;
+
+typedef struct thr_data {
+
+  int    threadno ;
+  int    NumBlocks ;
+  int    seed ;
+
+  int    min_size ;
+  int    max_size ;
+
+  char * *array ;
+  int    *blksize ;
+  int     asize ;
+
+  unsigned long    cAllocs ;
+  unsigned long    cFrees ;
+  int    cThreads ;
+  unsigned long    cBytesAlloced ;
+
+  volatile int finished ;
+  struct lran2_st rgen ;
+
+} thread_data;
+
+void runthreads(long sleep_cnt, int min_threads, int max_threads, 
+		int chperthread, int num_rounds) ;
+void runloops(long sleep_cnt, int num_chunks ) ;
+static void warmup(char **blkp, int num_chunks );
+static void * exercise_heap( void *pinput) ;
+static void lran2_init(struct lran2_st* d, long seed) ;
+static long lran2(struct lran2_st* d) ;
+ULONG CountReservedSpace() ;
+ 
+char **          blkp = new char *[MAX_BLOCKS] ;
+int *           blksize = new int[MAX_BLOCKS] ;
+long            seqlock=0 ;
+struct lran2_st rgen ;
+int             min_size=10, max_size=500 ;
+int             num_threads ;
+ULONG           init_space ;
+
+extern  int   cLockSleeps ;
+extern  int   cAllocedChunks ;
+extern  int   cAllocedSpace ;
+extern  int   cUsedSpace ;
+extern  int   cFreeChunks ;
+extern  int   cFreeSpace ;
+
+int cChecked=0 ;
+
+#if defined(_WIN32)
+extern "C" {
+  extern HANDLE crtheap;
+};
+#endif
+
+int main (int argc, char *argv[])
+{
+#if defined(USE_LFH) && defined(_WIN32)
+  // Activate 'Low Fragmentation Heap'.
+  ULONG info = 2;
+  HeapSetInformation (GetProcessHeap(),
+		      HeapCompatibilityInformation,
+		      &info,
+		      sizeof(info));
+#endif
+#if 0 // defined(__SVR4)
+ {
+   psinfo_t ps;
+   int pid = getpid();
+   char fname[255];
+   sprintf (fname, "/proc/%d/psinfo", pid);
+   // sprintf (fname, "/proc/self/ps");
+   FILE * f = fopen (fname, "rb");
+   printf ("opening %s\n", fname);
+   if (f) {
+     fread (&ps, sizeof(ps), 1, f);
+     printf ("resident set size = %dK\n", ps.pr_rssize);
+     fclose (f);
+   }
+ }
+#endif
+
+#if defined(_MT) || defined(_REENTRANT)
+  int          min_threads, max_threads ;
+  int          num_rounds ;
+  int          chperthread ;
+#endif
+  unsigned     seed=12345 ;
+  int          num_chunks=10000;
+  long sleep_cnt;
+
+  if (argc > 7) {
+    sleep_cnt = atoi(argv[1]);
+    min_size = atoi(argv[2]);
+    max_size = atoi(argv[3]);
+    chperthread = atoi(argv[4]);
+    num_rounds = atoi(argv[5]);
+    seed = atoi(argv[6]);
+    max_threads = atoi(argv[7]);
+    min_threads = max_threads;
+    printf ("sleep = %ld, min = %d, max = %d, per thread = %d, num rounds = %d, seed = %d, max_threads = %d, min_threads = %d\n",
+	    sleep_cnt, min_size, max_size, chperthread, num_rounds, seed, max_threads, min_threads);
+    goto DoneWithInput;
+  }
+
+#if defined(_MT) || defined(_REENTRANT)
+  //#ifdef _MT
+  printf( "\nMulti-threaded test driver \n") ;
+#else
+  printf( "\nSingle-threaded test driver \n") ;
+#endif
+#ifdef CPP
+  printf("C++ version (new and delete)\n") ;
+#else
+  printf("C version (malloc and free)\n") ;
+#endif
+  printf("runtime (sec): ") ;
+  scanf ("%ld", &sleep_cnt);
+
+  printf("chunk size (min,max): ") ;
+  scanf("%d %d", &min_size, &max_size ) ;
+#if defined(_MT) || defined(_REENTRANT)
+  //#ifdef _MT
+  printf("threads (min, max):   ") ; 
+  scanf("%d %d", &min_threads, &max_threads) ;
+  printf("chunks/thread:  ") ; scanf("%d", &chperthread ) ;
+  printf("no of rounds:   ") ; scanf("%d", &num_rounds ) ;
+  num_chunks = max_threads*chperthread ;
+#else 
+  printf("no of chunks:  ") ; scanf("%d", &num_chunks ) ;
+#endif
+  printf("random seed:    ") ; scanf("%d", &seed) ;
+
+ DoneWithInput:
+
+  if( num_chunks > MAX_BLOCKS ){
+    printf("Max %d chunks - exiting\n", MAX_BLOCKS ) ;
+    return(1) ;
+  }
+
+#ifndef __WIN32__
+#ifdef __SVR4
+  pthread_setconcurrency (max_threads);
+#endif
+#endif
+
+  lran2_init(&rgen, seed) ;
+  // init_space = CountReservedSpace() ;
+
+#if defined(_MT) || defined(_REENTRANT)
+  //#ifdef _MT
+  runthreads(sleep_cnt, min_threads, max_threads, chperthread, num_rounds) ;
+#else
+  runloops(sleep_cnt, num_chunks ) ;
+#endif
+
+#ifdef _DEBUG
+  _cputs("Hit any key to exit...") ;	(void)_getch() ;
+#endif
+
+  return 0;
+
+} /* main */
+
+void runloops(long sleep_cnt, int num_chunks )
+{
+  int     cblks ;
+  int     victim ;
+  int     blk_size ;
+#ifdef __WIN32__
+	_LARGE_INTEGER ticks_per_sec, start_cnt, end_cnt;
+#else
+  long ticks_per_sec ;
+  long start_cnt, end_cnt ;
+#endif
+  _int64        ticks ;
+  double        duration ;
+  double        reqd_space ;
+  ULONG         used_space ;
+  int           sum_allocs=0 ;
+
+  QueryPerformanceFrequency( &ticks_per_sec ) ;
+  QueryPerformanceCounter( &start_cnt) ;
+
+  for( cblks=0; cblks<num_chunks; cblks++){
+    if (max_size == min_size) {
+      blk_size = min_size;
+    } else {
+      blk_size = min_size+lran2(&rgen)%(max_size - min_size) ;
+    }
+#ifdef CPP
+    blkp[cblks] = new char[blk_size] ;
+#else
+    blkp[cblks] = (char *) malloc(blk_size) ;
+#endif
+    blksize[cblks] = blk_size ;
+    assert(blkp[cblks] != NULL) ;
+  }
+
+  while(TRUE){
+    for( cblks=0; cblks<num_chunks; cblks++){
+      victim = lran2(&rgen)%num_chunks ;
+#ifdef CPP
+      delete blkp[victim] ;
+#else
+      free(blkp[victim]) ;
+#endif
+
+      if (max_size == min_size) {
+	blk_size = min_size;
+      } else {
+	blk_size = min_size+lran2(&rgen)%(max_size - min_size) ;
+      }
+#ifdef CPP
+      blkp[victim] = new char[blk_size] ;
+#else
+      blkp[victim] = (char *) malloc(blk_size) ;
+#endif
+      blksize[victim] = blk_size ;
+      assert(blkp[victim] != NULL) ;
+    }
+    sum_allocs += num_chunks ;
+
+    QueryPerformanceCounter( &end_cnt) ;
+#ifdef __WIN32__
+		ticks = end_cnt.QuadPart - start_cnt.QuadPart ;
+		duration = (double)ticks/ticks_per_sec.QuadPart ;
+#else
+    ticks = end_cnt - start_cnt ;
+    duration = (double)ticks/ticks_per_sec ;
+#endif
+
+    if( duration >= sleep_cnt) break ;
+  }
+  reqd_space = (0.5*(min_size+max_size)*num_chunks) ;
+  // used_space = CountReservedSpace() - init_space;
+
+  printf("%6.3f", duration  ) ;
+  printf("%8.0f", sum_allocs/duration ) ;
+  printf(" %6.3f %.3f", (double)used_space/(1024*1024), used_space/reqd_space) ;
+  printf("\n") ;
+
+}
+
+
+#if defined(_MT) || defined(_REENTRANT)
+//#ifdef _MT
+void runthreads(long sleep_cnt, int min_threads, int max_threads, int chperthread, int num_rounds)
+{
+  thread_data *de_area = new thread_data[max_threads] ;
+  thread_data *pdea;
+  int           nperthread ;
+  int           sum_threads ;
+  unsigned long  sum_allocs ;
+  unsigned long  sum_frees ;
+  double        duration ;
+#ifdef __WIN32__
+	_LARGE_INTEGER ticks_per_sec, start_cnt, end_cnt;
+#else
+	long ticks_per_sec ;
+  long start_cnt, end_cnt ;
+#endif
+	_int64        ticks ;
+  double        rate_1=0, rate_n ;
+  double        reqd_space ;
+  ULONG         used_space ;
+  int           prevthreads ;
+  int           i ;
+
+  QueryPerformanceFrequency( &ticks_per_sec ) ;
+
+  pdea = &de_area[0] ;
+  memset(&de_area[0], 0, sizeof(thread_data)) ;
+
+  prevthreads = 0 ;
+  for(num_threads=min_threads; num_threads <= max_threads; num_threads++ )
+    {
+
+      warmup(&blkp[prevthreads*chperthread], (num_threads-prevthreads)*chperthread );
+
+      nperthread = chperthread ;
+      stopflag   = FALSE ;
+		
+      for(i=0; i< num_threads; i++){
+	de_area[i].threadno    = i+1 ;
+	de_area[i].NumBlocks   = num_rounds*nperthread;
+	de_area[i].array       = &blkp[i*nperthread] ;
+	de_area[i].blksize     = &blksize[i*nperthread] ;
+	de_area[i].asize       = nperthread ;
+	de_area[i].min_size    = min_size ;
+	de_area[i].max_size    = max_size ;
+	de_area[i].seed        = lran2(&rgen) ; ;
+	de_area[i].finished    = 0 ;
+	de_area[i].cAllocs     = 0 ;
+	de_area[i].cFrees      = 0 ;
+	de_area[i].cThreads    = 0 ;
+	de_area[i].finished    = FALSE ;
+	lran2_init(&de_area[i].rgen, de_area[i].seed) ;
+
+#ifdef __WIN32__
+	_beginthread((void (__cdecl*)(void *)) exercise_heap, 0, &de_area[i]) ;  
+#else
+	_beginthread(exercise_heap, 0, &de_area[i]) ;  
+#endif
+
+	}
+
+      QueryPerformanceCounter( &start_cnt) ;
+
+      // printf ("Sleeping for %ld seconds.\n", sleep_cnt);
+      Sleep(sleep_cnt * 1000L) ;
+
+      stopflag = TRUE ;
+
+      for(i=0; i<num_threads; i++){
+	while( !de_area[i].finished ){
+#ifdef __WIN32__
+		Sleep(1);
+#elif defined(__SVR4)
+		thr_yield();
+#else
+		sched_yield();
+#endif
+	}
+      }
+
+
+      QueryPerformanceCounter( &end_cnt) ;
+
+      sum_frees = sum_allocs =0  ;
+      sum_threads = 0 ;
+      for(i=0;i< num_threads; i++){
+	sum_allocs    += de_area[i].cAllocs ;
+	sum_frees     += de_area[i].cFrees ;
+	sum_threads   += de_area[i].cThreads ;
+	de_area[i].cAllocs = de_area[i].cFrees = 0;
+      }
+
+ 
+#ifdef __WIN32__
+      ticks = end_cnt.QuadPart - start_cnt.QuadPart ;
+     duration = (double)ticks/ticks_per_sec.QuadPart ;
+#else
+      ticks = end_cnt - start_cnt ;
+     duration = (double)ticks/ticks_per_sec ;
+#endif
+
+      for( i=0; i<num_threads; i++){
+	if( !de_area[i].finished )
+	  printf("Thread at %d not finished\n", i) ;
+      }
+
+
+      rate_n = sum_allocs/duration ;
+      if( rate_1 == 0){
+	rate_1 = rate_n ;
+      }
+		
+      reqd_space = (0.5*(min_size+max_size)*num_threads*chperthread) ;
+      // used_space = CountReservedSpace() - init_space;
+      used_space = 0;
+      
+      printf ("Throughput = %8.0f operations per second.\n", sum_allocs / duration);
+
+#if 0
+      printf("%2d ", num_threads ) ;
+      printf("%6.3f", duration  ) ;
+      printf("%6.3f", rate_n/rate_1 ) ;
+      printf("%8.0f", sum_allocs/duration ) ;
+      printf(" %6.3f %.3f", (double)used_space/(1024*1024), used_space/reqd_space) ;
+      printf("\n") ;
+#endif
+
+      Sleep(5000L) ; // wait 5 sec for old threads to die
+
+      prevthreads = num_threads ;
+
+      printf ("Done sleeping...\n");
+
+    }
+  delete [] de_area;
+}
+
+
+static void * exercise_heap( void *pinput)
+{
+  thread_data  *pdea;
+  int           cblks=0 ;
+  int           victim ;
+  long          blk_size ;
+  int           range ;
+
+  if( stopflag ) return 0;
+
+  pdea = (thread_data *)pinput ;
+  pdea->finished = FALSE ;
+  pdea->cThreads++ ;
+  range = pdea->max_size - pdea->min_size ;
+
+  /* allocate NumBlocks chunks of random size */
+  for( cblks=0; cblks<pdea->NumBlocks; cblks++){
+    victim = lran2(&pdea->rgen)%pdea->asize ;
+#ifdef CPP
+    delete pdea->array[victim] ;
+#else
+    free(pdea->array[victim]) ;
+#endif
+    pdea->cFrees++ ;
+
+    if (range == 0) {
+      blk_size = pdea->min_size;
+    } else {
+      blk_size = pdea->min_size+lran2(&pdea->rgen)%range ;
+    }
+#ifdef CPP
+    pdea->array[victim] = new char[blk_size] ;
+#else
+    pdea->array[victim] = (char *) malloc(blk_size) ;
+#endif
+
+    pdea->blksize[victim] = blk_size ;
+    assert(pdea->array[victim] != NULL) ;
+
+    pdea->cAllocs++ ;
+
+		/* Write something! */
+
+		volatile char * chptr = ((char *) pdea->array[victim]);
+		*chptr++ = 'a';
+		volatile char ch = *((char *) pdea->array[victim]);
+		*chptr = 'b';
+
+    
+		if( stopflag ) break ;
+  }
+
+  //  	printf("Thread %u terminating: %d allocs, %d frees\n",
+  //		      pdea->threadno, pdea->cAllocs, pdea->cFrees) ;
+  pdea->finished = TRUE ;
+
+  if( !stopflag ){
+#ifdef __WIN32__
+	_beginthread((void (__cdecl*)(void *)) exercise_heap, 0, pdea) ;  
+#else
+    _beginthread(exercise_heap, 0, pdea) ;
+#endif
+  } else {
+    printf ("thread stopping.\n");
+  }
+#ifndef _WIN32
+  pthread_exit (NULL);
+#endif
+  return 0;
+}
+
+static void warmup(char **blkp, int num_chunks )
+{
+  int     cblks ;
+  int     victim ;
+  int     blk_size ;
+  LPVOID  tmp ;
+
+
+  for( cblks=0; cblks<num_chunks; cblks++){
+    if (min_size == max_size) {
+      blk_size = min_size;
+    } else {
+      blk_size = min_size+lran2(&rgen)%(max_size-min_size) ;
+    }
+#ifdef CPP
+    blkp[cblks] = new char[blk_size] ;
+#else
+    blkp[cblks] = (char *) malloc(blk_size) ;
+#endif
+    blksize[cblks] = blk_size ;
+    assert(blkp[cblks] != NULL) ;
+  }
+
+  /* generate a random permutation of the chunks */
+  for( cblks=num_chunks; cblks > 0 ; cblks--){
+    victim = lran2(&rgen)%cblks ;
+    tmp = blkp[victim] ;
+    blkp[victim]  = blkp[cblks-1] ;
+    blkp[cblks-1] = (char *) tmp ;
+  }
+
+  for( cblks=0; cblks<4*num_chunks; cblks++){
+    victim = lran2(&rgen)%num_chunks ;
+#ifdef CPP
+    delete blkp[victim] ;
+#else
+    free(blkp[victim]) ;
+#endif
+
+    if (max_size == min_size) {
+      blk_size = min_size;
+    } else {
+      blk_size = min_size+lran2(&rgen)%(max_size - min_size) ;
+    }
+#ifdef CPP
+    blkp[victim] = new char[blk_size] ;
+#else
+    blkp[victim] = (char *) malloc(blk_size) ;
+#endif
+    blksize[victim] = blk_size ;
+    assert(blkp[victim] != NULL) ;
+  }
+}
+#endif // _MT
+
+#ifdef __WIN32__
+ULONG CountReservedSpace()
+{
+  MEMORY_BASIC_INFORMATION info;
+  char                     *addr=NULL ;
+  ULONG                     size=0 ;
+
+  while( true){
+    VirtualQuery(addr, &info, sizeof(info));
+    switch( info.State){
+    case MEM_FREE:
+    case MEM_RESERVE:
+      break ;
+    case MEM_COMMIT:
+      size += info.RegionSize ;
+      break ;
+    }
+    addr += info.RegionSize ;
+    if( addr >= (char *)0x80000000UL ) break ;
+  }
+
+  return size ;
+
+}
+#endif
+
+// =======================================================
+
+/* lran2.h
+ * by Wolfram Gloger 1996.
+ *
+ * A small, portable pseudo-random number generator.
+ */
+
+#ifndef _LRAN2_H
+#define _LRAN2_H
+
+#define LRAN2_MAX 714025l /* constants for portable */
+#define IA	  1366l	  /* random number generator */
+#define IC	  150889l /* (see e.g. `Numerical Recipes') */
+
+//struct lran2_st {
+//    long x, y, v[97];
+//};
+
+static void
+lran2_init(struct lran2_st* d, long seed)
+{
+  long x;
+  int j;
+
+  x = (IC - seed) % LRAN2_MAX;
+  if(x < 0) x = -x;
+  for(j=0; j<97; j++) {
+    x = (IA*x + IC) % LRAN2_MAX;
+    d->v[j] = x;
+  }
+  d->x = (IA*x + IC) % LRAN2_MAX;
+  d->y = d->x;
+}
+
+static 
+long lran2(struct lran2_st* d)
+{
+  int j = (d->y % 97);
+
+  d->y = d->v[j];
+  d->x = (IA*d->x + IC) % LRAN2_MAX;
+  d->v[j] = d->x;
+  return d->y;
+}
+
+#undef IA
+#undef IC
+
+#endif
+
+
diff --git a/src/benchmarks/timer.h b/src/benchmarks/timer.h
new file mode 100644
index 0000000..d4d42c7
--- /dev/null
+++ b/src/benchmarks/timer.h
@@ -0,0 +1,372 @@
+/* -*- C++ -*- */
+
+/*
+
+  Heap Layers: An Extensible Memory Allocation Infrastructure
+  
+  Copyright (C) 2000-2003 by Emery Berger
+  http://www.cs.umass.edu/~emery
+  emery@cs.umass.edu
+  
+  This program is free software; you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation; either version 2 of the License, or
+  (at your option) any later version.
+  
+  This program is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+  
+  You should have received a copy of the GNU General Public License
+  along with this program; if not, write to the Free Software
+  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+
+*/
+
+#include <cassert>
+#include <stdio.h>
+
+
+#ifndef _TIMER_H_
+#define _TIMER_H_
+
+/**
+ * @class Timer
+ * @brief A portable class for high-resolution timing.
+ *
+ * This class simplifies timing measurements across a number of platforms.
+ * 
+ * @code
+ *  Timer t;
+ *  t.start();
+ *  // do some work
+ *  t.stop();
+ *  cout << "That took " << (double) t << " seconds." << endl;
+ * @endcode
+ *
+ */
+
+#ifdef __APPLE__
+#include <sys/time.h>
+#endif
+
+#if defined(__linux__) && defined(__GNUG__) && defined(__i386__)
+
+#include <stdio.h>
+#include <limits.h>
+#include <time.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <string.h>
+
+static void getTime (unsigned long& tlo, unsigned long& thi) {
+  asm volatile ("rdtsc"
+		: "=a"(tlo),
+		"=d" (thi));
+}
+
+
+static double getFrequency (void) {
+  static double freq = 0.0;
+  static bool initialized = false;
+  unsigned long LTime0, LTime1, HTime0, HTime1;
+  if (!initialized) { 
+
+    freq = 2600000.0;
+
+#if 0
+    // Compute MHz directly.
+    // Wait for approximately one second.
+    
+    getTime (LTime0, HTime0);
+    //    printf ("waiting...\n");
+    struct timespec rqtp, rmtp;
+    rqtp.tv_sec = 1;
+    rqtp.tv_nsec = 0;
+    nanosleep (&rqtp, &rmtp);
+    // printf ("done.\n");
+    getTime (LTime1, HTime1);
+
+    freq = (double)(LTime1 - LTime0) + (double)(UINT_MAX)*(double)(HTime1 - HTime0);
+    if (LTime1 < LTime0) {
+      freq -= (double)UINT_MAX;
+    }
+#endif
+    initialized = true;
+
+  } else {
+    // printf ("wha?\n");
+  }
+  return freq;
+}
+
+
+namespace HL {
+
+class Timer {
+public:
+  Timer (void)
+    : timeElapsed (0.0)
+  {
+    _frequency = getFrequency();
+    //    printf ("wooo!\n");
+    //  printf ("freq = %lf\n", frequency);
+  }
+  void start (void) {
+    getTime (currentLo, currentHi);
+  }
+  void stop (void) {
+    unsigned long lo, hi;
+    getTime (lo, hi);
+    double now = (double) hi * 4294967296.0 + lo;
+    double prev = (double) currentHi * 4294967296.0 + currentLo;
+    timeElapsed = (now - prev) / _frequency;
+  }
+
+  operator double (void) {
+    return timeElapsed;
+  }
+
+private:
+  double timeElapsed;
+  unsigned long currentLo, currentHi;
+  double _frequency;
+};
+
+};
+
+#else
+
+
+#ifdef __SVR4 // Solaris
+#include <sys/time.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <sys/procfs.h>
+#include <stdio.h>
+#endif // __SVR4
+
+#include <time.h>
+
+#if defined(unix) || defined(__linux)
+#include <sys/time.h>
+#include <unistd.h>
+#endif
+
+
+#ifdef __sgi
+#include <sys/types.h>
+#include <sys/times.h>
+#include <limits.h>
+#endif
+
+
+#if defined(_WIN32)
+#include <windows.h>
+#endif
+
+
+#if defined(__BEOS__)
+#include <OS.h>
+#endif
+
+
+namespace HL {
+
+class Timer {
+
+public:
+
+  /// Initializes the timer.
+  Timer (void)
+#if !defined(_WIN32)
+    : _starttime (0),
+      _elapsedtime (0)
+#endif
+  {
+  }
+
+  /// Start the timer.
+  void start (void) { _starttime = _time(); }
+
+  /// Stop the timer.
+  void stop (void) { _elapsedtime += _time() - _starttime; }
+
+  /// Reset the timer.
+  void reset (void) { _starttime = _elapsedtime; }
+
+#if 0
+  // Set the timer.
+  void set (double secs) { _starttime = 0; _elapsedtime = _sectotime (secs);}
+#endif
+
+  /// Return the number of seconds elapsed.
+  operator double (void) { return _timetosec (_elapsedtime); }
+
+  static double currentTime (void) { TimeType t; t = _time(); return _timetosec (t); }
+
+
+private:
+
+  // The _timer variable will be different depending on the OS.
+  // We try to use the best timer available.
+
+#ifdef __sgi
+#define TIMER_FOUND
+
+  long _starttime, _elapsedtime;
+
+  long _time (void) {
+    struct tms t;
+    long ticks = times (&t);
+    return ticks;
+  }
+
+  static double _timetosec (long t) {
+    return ((double) (t) / CLK_TCK);
+  }
+
+  static long _sectotime (double sec) {
+    return (long) sec * CLK_TCK;
+  }
+#endif
+
+#ifdef __SVR4 // Solaris
+#define TIMER_FOUND
+  typedef hrtime_t TimeType;
+  TimeType	_starttime, _elapsedtime;
+
+  static TimeType _time (void) {
+    return gethrtime();
+  }
+
+  static TimeType _sectotime (double sec) { return (hrtime_t) (sec * 1.0e9); }
+
+  static double _timetosec (TimeType& t) {
+    return ((double) (t) / 1.0e9);
+  }
+#endif // __SVR4
+
+#if defined(MAC) || defined(macintosh)
+#define TIMER_FOUND
+  double		_starttime, _elapsedtime;
+
+  double _time (void) {
+    return get_Mac_microseconds();
+  }
+
+  double _timetosec (hrtime_t& t) {
+    return t;
+  }
+#endif // MAC
+
+#ifdef _WIN32
+#define TIMER_FOUND
+
+#ifndef __GNUC__
+  class TimeType {
+  public:
+    TimeType (void)
+    {
+      largeInt.QuadPart = 0;
+    }
+    operator double& (void) { return (double&) largeInt.QuadPart; }
+    operator LARGE_INTEGER& (void) { return largeInt; }
+    double timeToSec (void) {
+      return (double) largeInt.QuadPart / getFreq();
+    }
+  private:
+    double getFreq (void) {
+      QueryPerformanceFrequency (&freq);
+      return (double) freq.QuadPart;
+    }
+
+    LARGE_INTEGER largeInt;
+    LARGE_INTEGER freq;
+  };
+
+  TimeType _starttime, _elapsedtime;
+
+  static TimeType _time (void) {
+    TimeType t;
+    int r = QueryPerformanceCounter (&((LARGE_INTEGER&) t));
+    assert (r);
+    return t;
+  }
+
+  static double _timetosec (TimeType& t) {
+    return t.timeToSec();
+  }
+#else
+  typedef DWORD TimeType;
+  DWORD _starttime, _elapsedtime;
+  static DWORD _time (void) {
+    return GetTickCount();
+  }
+
+  static double _timetosec (DWORD& t) {
+    return (double) t / 100000.0;
+  }
+  static unsigned long _sectotime (double sec) {
+    return (unsigned long)(sec);
+  }
+#endif
+#endif // _WIN32
+
+
+#ifdef __BEOS__
+#define TIMER_FOUND
+  bigtime_t _starttime, _elapsedtime;
+  bigtime_t _time(void) {
+    return system_time();
+  }
+  double _timetosec (bigtime_t& t) {
+    return (double) t / 1000000.0;
+  }
+  
+  bigtime_t _sectotime (double sec) {
+    return (bigtime_t)(sec * 1000000.0);
+  }
+#endif // __BEOS__
+
+#ifndef TIMER_FOUND
+
+  typedef long TimeType;
+  TimeType _starttime, _elapsedtime;
+
+  static TimeType _time (void) {
+    struct timeval t;
+    gettimeofday (&t, NULL);
+    return t.tv_sec * 1000000 + t.tv_usec;
+  }
+
+  static double _timetosec (TimeType t) {
+    return ((double) (t) / 1000000.0);
+  }
+
+  static TimeType _sectotime (double sec) {
+    return (TimeType) (sec * 1000000.0);
+  }
+
+#endif // TIMER_FOUND
+
+#undef TIMER_FOUND
+
+};
+
+
+#ifdef __SVR4 // Solaris
+class VirtualTimer : public Timer {
+public:
+  hrtime_t _time (void) {
+    return gethrvtime();
+  }
+};  
+#endif
+
+}
+
+#endif
+
+#endif
diff --git a/src/chattymalloc.c b/src/chattymalloc.c
new file mode 100644
index 0000000..54708d6
--- /dev/null
+++ b/src/chattymalloc.c
@@ -0,0 +1,161 @@
+#define _GNU_SOURCE
+#include <dlfcn.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <stdarg.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+static char tmpbuff[1024];
+static unsigned long tmppos = 0;
+static unsigned long tmpallocs = 0;
+
+static int out = -1;
+static int prevent_recursion = 0;
+
+/*=========================================================
+ *  * interception points
+ *  */
+
+static void * (*myfn_malloc)(size_t size);
+static void (*myfn_free)(void* ptr);
+static void * (*myfn_calloc)(size_t nmemb, size_t size);
+static void * (*myfn_realloc)(void* ptr, size_t size);
+static void * (*myfn_memalign)(size_t alignment, size_t size);
+
+static void write_output(const char* fmt, ...)
+{
+	if (!prevent_recursion)
+	{
+		prevent_recursion = 1;
+
+		/* lockf(out, F_LOCK, 0); */
+
+		va_list args;
+		va_start(args, fmt);
+		vdprintf(out, fmt, args);
+		va_end(args);
+
+		/* lockf(out, F_ULOCK, 0); */
+		prevent_recursion = 0;
+	}
+}
+
+static void init()
+{
+	out = open("chattymalloc.data", O_WRONLY | O_TRUNC | O_CREAT, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
+	if (out == -1)
+	{
+		fprintf(stderr, "failed to open output file with %d\n", errno);
+		exit(1);
+	}
+
+	myfn_malloc     = dlsym(RTLD_NEXT, "malloc");
+	myfn_free       = dlsym(RTLD_NEXT, "free");
+	myfn_calloc     = dlsym(RTLD_NEXT, "calloc");
+	myfn_realloc    = dlsym(RTLD_NEXT, "realloc");
+	myfn_memalign   = dlsym(RTLD_NEXT, "memalign");
+
+	if (!myfn_malloc || !myfn_free || !myfn_calloc || !myfn_realloc || !myfn_memalign)
+	{
+		fprintf(stderr, "Error in `dlsym`: %s\n", dlerror());
+		exit(1);
+	}
+}
+
+void *malloc(size_t size)
+{
+	static int initializing = 0;
+
+	if (myfn_malloc == NULL)
+	{
+		if (!initializing)
+		{
+			initializing = 1;
+			init();
+			initializing = 0;
+
+		}
+		else
+		{
+			if (tmppos + size < sizeof(tmpbuff))
+			{
+				void *retptr = tmpbuff + tmppos;
+				tmppos += size;
+				++tmpallocs;
+				return retptr;
+			}
+			else
+			{
+				fprintf(stderr, "%d in %d allocs\n", tmppos, tmpallocs);
+				fprintf(stderr, "jcheck: too much memory requested during initialisation - increase tmpbuff size\n");
+				exit(1);
+			}
+		}
+	}
+
+	void *ptr = myfn_malloc(size);
+	write_output("m %zu %p\n", size, ptr);
+	return ptr;
+}
+
+void free(void *ptr)
+{
+	// something wrong if we call free before one of the allocators!
+	if (myfn_malloc == NULL)
+		init();
+	if (!(ptr >= (void*) tmpbuff && ptr <= (void*)(tmpbuff + tmppos)))
+	{
+		write_output("f %p\n", ptr);
+		myfn_free(ptr);
+	}
+}
+
+void* realloc(void *ptr, size_t size)
+{
+	if (myfn_realloc == NULL)
+	{
+		void *nptr = malloc(size);
+		if (nptr && ptr)
+		{
+			memmove(nptr, ptr, size);
+			free(ptr);
+		}
+		return nptr;
+	}
+
+	void* nptr =  myfn_realloc(ptr, size);
+	write_output("r %p %zu %p\n", ptr, size, nptr);
+	return nptr;
+}
+
+void* calloc(size_t nmemb, size_t size)
+{
+	if (myfn_calloc == NULL)
+	{
+		void *ptr = malloc(nmemb*size);
+		if (ptr)
+			memset(ptr, 0, nmemb*size);
+		return ptr;
+	}
+
+	void* ptr = myfn_calloc(nmemb, size);
+	write_output("c %zu %zu %p\n", nmemb, size, ptr);
+	return ptr;
+}
+
+void* memalign(size_t alignment, size_t size)
+{
+	if (myfn_memalign == NULL)
+	{
+		fprintf(stderr, "called memalign before or during init");
+		exit(1);
+	}
+
+	void* ptr = myfn_memalign(alignment, size);
+	write_output("mm %zu %zu %p\n", alignment, size, ptr);
+	return ptr;
+}
diff --git a/src/chattyparser.py b/src/chattyparser.py
new file mode 100644
index 0000000..3406b44
--- /dev/null
+++ b/src/chattyparser.py
@@ -0,0 +1,158 @@
+import re
+import matplotlib.pyplot as plt
+import numpy as np
+
+ptr = "(?:0x)?(?P<ptr>(?:\w+)|(?:\(nil\)))"
+size = "(?P<size>\d+)"
+
+malloc_re = re.compile("^m {} {}$".format(size, ptr))
+free_re = re.compile("^f {}$".format(ptr))
+calloc_re = re.compile("^c (?P<nmemb>\d+) {} {}$".format(size, ptr))
+realloc_re = re.compile("^r {} {} {}$".format(ptr, size, ptr.replace("ptr", "nptr")))
+memalign_re = re.compile("^mm (?P<alignment>\d+) {} {}$".format(size, ptr))
+
+def record_allocation(hist, total_size, allocations, ptr, size, coll_size, req_size, nohist, optr=None, add=True):
+    size = int(size)
+    if add:
+        if optr and optr in allocations:
+            size -= allocations[optr]
+            del(allocations[optr])
+
+        allocations[ptr] = size
+        if not nohist:
+            hist[size] = hist.get(size, 0) + 1
+
+        if type(total_size[-1]) != int or type(size) != int:
+            print("invalid type", type(total_size[-1]), type(size))
+            return
+
+        if coll_size:
+            if not req_size or size == req_size:
+                total_size.append(total_size[-1] + size)
+            elif req_size:
+                total_size.append(total_size[-1])
+
+    elif ptr != "(nil)" and ptr in allocations:
+        size = allocations[ptr]
+        if coll_size:
+            if not req_size or size == req_size:
+                total_size.append(total_size[-1] - size)
+            elif req_size:
+                total_size.append(total_size[-1])
+
+        del(allocations[ptr])
+    elif coll_size:
+        total_size.append(total_size[-1])
+
+def parse(path="chattymalloc.data", coll_size=True, req_size=None, nohist=False):
+    tmalloc, tcalloc, trealloc, tfree, tmemalign= 0, 0, 0, 0, 0
+    allocations = {}
+    requested_size = [0]
+    hist = {}
+    ln = 0
+
+    with open(path, "r") as f:
+        for i, l in enumerate(f.readlines()):
+            ln += 1
+            res = malloc_re.match(l)
+            if res != None:
+                res = res.groupdict()
+                record_allocation(hist, requested_size, allocations, res["ptr"],
+                    res["size"], coll_size, req_size, nohist)
+                tmalloc += 1
+                continue
+
+            res = free_re.match(l)
+            if res != None:
+                res = res.groupdict()
+                record_allocation(hist, requested_size, allocations, res["ptr"],
+                    0, coll_size, req_size, nohist, add=False)
+                tfree +=1
+                continue
+
+            res = calloc_re.match(l)
+            if res != None:
+                res = res.groupdict()
+                size = int(res["nmemb"]) * int(res["size"])
+                record_allocation(hist, requested_size, allocations, res["ptr"],
+                    size, coll_size, req_size, nohist)
+                tcalloc += 1
+                continue
+
+            res = realloc_re.match(l)
+            if res != None:
+                res = res.groupdict()
+                record_allocation(hist, requested_size, allocations, res["nptr"],
+                    res["size"], coll_size, req_size, nohist, optr=res["ptr"])
+                trealloc += 1
+                continue
+
+            res = memalign_re.match(l)
+            if res != None:
+                res = res.groupdict()
+                record_allocation(hist, requested_size, allocations, res["ptr"],
+                    res["size"], coll_size, req_size, nohist)
+                tmemalign += 1
+                continue
+
+            print("\ninvalid line at", ln, ":", l)
+    calls = {"malloc": tmalloc, "free": tfree, "calloc": tcalloc, "realloc": trealloc, "memalign": tmemalign}
+    return hist, calls, requested_size
+
+def plot(path):
+    hist, calls, _ = parse(req_size=None)
+    plot_hist_ascii(path+".hist", hist, calls)
+    top5 = [t[1] for t in sorted([(n, s) for s, n in hist.items()])[-5:]]
+
+    del(hist)
+    del(calls)
+    plot_profile(path+".profile.png", top5)
+
+
+def plot_profile(path, top5):
+    _, calls, total_size = parse(nohist=True)
+    x_vals = range(0, sum(calls.values()) + 1)
+
+    plt.plot(x_vals, total_size, marker='', linestyle='-', label="Total requested")
+
+    for s in top5:
+        _, calls, total_size = parse(nohist=True, req_size=s)
+        plt.plot(x_vals, total_size, label=s)
+
+    plt.legend()
+    plt.xlabel("Allocations")
+    plt.ylabel("mem in kb")
+    plt.title("Memusage profile")
+    plt.savefig(path)
+    plt.clf()
+
+def plot_hist_ascii(path, hist, calls):
+    bins = {}
+    for size in sorted(hist):
+        bin = int(size / 16)
+        bins[bin] = bins.get(bin, 0) + hist[size]
+
+    total = sum(calls.values()) - calls["free"]
+    with open(path, "w") as f:
+        print("Total function calls:", total, file=f)
+        print("malloc:", calls["malloc"], file=f)
+        print("calloc:", calls["calloc"], file=f)
+        print("realloc:", calls["realloc"], file=f)
+        print("free:", calls["free"], file=f)
+        print("memalign:", calls["memalign"], file=f)
+        print(file=f)
+
+        print("< 1024", sum([n for s,n in hist.items() if s < 1024]), file=f)
+        print("< 4096", sum([n for s,n in hist.items() if s < 4096]), file=f)
+        print(file=f)
+
+        print("Histogram of sizes:", file=f)
+        sbins = sorted(bins)
+        binmaxlength = str(len(str(sbins[-1])) + 1)
+        amountmaxlength = str(len(str(sorted(bins.values())[-1])))
+        for b in sbins:
+            perc = bins[b]/total*100
+            binsize = "{:<" + binmaxlength + "} - {:>" + binmaxlength + "}"
+            print(binsize.format((b)*16, (b+1)*16-1), end=" ", file=f)
+            amount = "{:<" + amountmaxlength + "} {:.2f}% {}"
+            print(amount.format(bins[b], perc, '*'*int(perc/2)), file=f)
diff --git a/src/larson.py b/src/larson.py
new file mode 100644
index 0000000..713818b
--- /dev/null
+++ b/src/larson.py
@@ -0,0 +1,54 @@
+import multiprocessing
+import re
+
+from src.benchmark import Benchmark
+
+throughput_re = re.compile("^Throughput =\s*(?P<throughput>\d+) operations per second.$")
+
+class Benchmark_Larson( Benchmark ):
+    def __init__(self):
+        self.name = "larson"
+        self.descrition = """This benchmark is courtesy of Paul Larson at Microsoft
+                             Research. It simulates a server: each thread allocates
+                             and deallocates objects, and then transfers some objects
+                             (randomly selected) to other threads to be freed."""
+
+        self.cmd = "build/larson{binary_suffix} 1 8 {maxsize} 1000 50000 1 {threads}"
+
+        self.args = {
+                        "maxsize" : [8, 32, 64, 128, 256, 512, 1024],
+                        "threads" : range(1, multiprocessing.cpu_count() * 2 + 1)
+                    }
+
+        self.requirements = ["build/larson"]
+        super().__init__()
+
+    def process_output(self, result, stdout, stderr, target, perm, verbose):
+        for l in stdout.splitlines():
+            res = throughput_re.match(l)
+            if res:
+                result["throughput"] = int(res.group("throughput"))
+                return
+
+    def summary(self, sumdir):
+        # Plot threads->throughput and maxsize->throughput
+        self.plot_fixed_arg("{throughput}/1000000",
+                    ylabel="'MOPS/s'",
+                    title = "'Larson: ' + arg + ' ' + str(arg_value)",
+                    filepostfix = "throughput",
+                    sumdir=sumdir)
+
+        self.plot_fixed_arg("({L1-dcache-load-misses}/{L1-dcache-loads})*100",
+                    ylabel="'l1 cache misses in %'",
+                    title = "'Larson cache misses: ' + arg + ' ' + str(arg_value)",
+                    filepostfix = "cachemisses",
+                    sumdir=sumdir)
+
+        # Memusage
+        self.plot_fixed_arg("int({VmHWM})",
+                    ylabel='"VmHWM in kB"',
+                    title= '"Loop Memusage: " + arg + " " + str(arg_value)',
+                    filepostfix="memusage",
+                    sumdir=sumdir)
+
+larson = Benchmark_Larson()
diff --git a/src/print_status_on_exit.c b/src/print_status_on_exit.c
new file mode 100644
index 0000000..1d72384
--- /dev/null
+++ b/src/print_status_on_exit.c
@@ -0,0 +1,37 @@
+#define _GNU_SOURCE
+#include <dlfcn.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+static void print_status(void)
+{
+	char buf[4096];
+
+	FILE* status = fopen("/proc/self/status", "r");
+	if (status == NULL)
+	{
+		perror("fopen status");
+		exit(1);
+	}
+
+	FILE* output = fopen("status", "a");
+	if (output == NULL)
+	{
+		perror("fopen output file");
+		exit(1);
+	}
+
+	while (!feof(status))
+	{
+		fgets(&buf, 4096, status);
+		fprintf(output, "%s", buf);
+	}
+	fclose(status);
+}
+
+static void __attribute__((constructor)) init()
+{
+	atexit(print_status);
+}
+
diff --git a/src/trace_run.c b/src/trace_run.c
new file mode 100644
index 0000000..604d01e
--- /dev/null
+++ b/src/trace_run.c
@@ -0,0 +1,750 @@
+#define _LARGEFILE64_SOURCE
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <pthread.h>
+#include <sys/time.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/resource.h>
+#include <fcntl.h>
+#include <unistd.h>
+
+// #include "malloc.h"
+#include <malloc.h>
+
+// #include "mtrace.h"
+/* Codes for the simulator/workload programs. Copied from mtrace.h. */
+#define C_NOP 0
+#define C_DONE 1
+#define C_MALLOC 2
+#define C_CALLOC 3
+#define C_REALLOC 4
+#define C_FREE 5
+#define C_SYNC_W 6
+#define C_SYNC_R 7
+#define C_ALLOC_PTRS 8
+#define C_ALLOC_SYNCS 9
+#define C_NTHREADS 10
+#define C_START_THREAD 11
+#define C_MEMALIGN 12
+#define C_VALLOC 13
+#define C_PVALLOC 14
+#define C_POSIX_MEMALIGN 15
+
+#if UINTPTR_MAX == 0xffffffffffffffff
+
+#define ticks_t int64_t
+/* Setting quick_run to 1 allows the simulator to model
+   only the allocation and deallocation accounting via
+   atomic_rss. The actual allocations are skipped.  This
+   mode is useful to verify the workload file.  */
+#define quick_run 0
+
+static __inline__ ticks_t rdtsc_s(void)
+{
+  unsigned a, d;
+  asm volatile("cpuid" ::: "%rax", "%rbx", "%rcx", "%rdx");
+  asm volatile("rdtscp" : "=a" (a), "=d" (d));
+  return ((unsigned long long)a) | (((unsigned long long)d) << 32);
+}
+
+static __inline__ ticks_t rdtsc_e(void)
+{
+  unsigned a, d;
+  asm volatile("rdtscp" : "=a" (a), "=d" (d));
+  asm volatile("cpuid" ::: "%rax", "%rbx", "%rcx", "%rdx");
+  return ((unsigned long long)a) | (((unsigned long long)d) << 32);
+}
+
+#else
+
+#define ticks_t int32_t
+
+static __inline__ ticks_t rdtsc_s(void)
+{
+  unsigned a, d;
+  asm volatile("cpuid" ::: "%ax", "%bx", "%cx", "%dx");
+  asm volatile("rdtsc" : "=a" (a), "=d" (d));
+  return ((unsigned long)a) | (((unsigned long)d) << 16);
+}
+
+static __inline__ ticks_t rdtsc_e(void)
+{
+  unsigned a, d;
+  asm volatile("rdtscp" : "=a" (a), "=d" (d));
+  asm volatile("cpuid" ::: "%ax", "%bx", "%cx", "%dx");
+  return ((unsigned long)a) | (((unsigned long)d) << 16);
+}
+
+#endif
+
+static ticks_t diff_timeval (struct timeval e, struct timeval s)
+{
+  ticks_t usec;
+  if (e.tv_usec < s.tv_usec)
+    usec = (e.tv_usec + 1000000 - s.tv_usec) + (e.tv_sec-1 - s.tv_sec)*1000000;
+  else
+    usec = (e.tv_usec - s.tv_usec) + (e.tv_sec - s.tv_sec)*1000000;
+  return usec;
+}
+
+#if 1
+#define Q1
+#define Q2
+#else
+pthread_mutex_t genmutex = PTHREAD_MUTEX_INITIALIZER;
+#define Q1   pthread_mutex_lock(&genmutex)
+#define Q2   pthread_mutex_unlock(&genmutex)
+#endif
+
+pthread_mutex_t cmutex = PTHREAD_MUTEX_INITIALIZER;
+#define NCBUF 10
+static char cbuf[NCBUF][30];
+static int ci = 0;
+
+char *comma(ticks_t x)
+{
+  char buf[30], *bs, *bd;
+  int l, i, idx;
+
+  pthread_mutex_lock(&cmutex);
+  ci = (ci + 1) % NCBUF;
+  idx = ci;
+  pthread_mutex_unlock(&cmutex);
+  bs = buf;
+  bd = cbuf[idx];
+
+  sprintf(buf, "%lld", (long long int)x);
+  l = strlen(buf);
+  i = l;
+  while (*bs)
+    {
+      *bd++ = *bs++;
+      i--;
+      if (i % 3 == 0 && *bs)
+	*bd++ = ',';
+    }
+  *bd = 0;
+  return cbuf[idx];
+}
+
+static volatile void **ptrs;
+static volatile size_t *sizes;
+static size_t n_ptrs;
+static volatile char *syncs;
+static pthread_mutex_t *mutexes;
+static pthread_cond_t *conds;
+static size_t n_syncs;
+
+static pthread_mutex_t stat_mutex = PTHREAD_MUTEX_INITIALIZER;
+ticks_t malloc_time = 0, malloc_count = 0;
+ticks_t calloc_time = 0, calloc_count = 0;
+ticks_t realloc_time = 0, realloc_count = 0;
+ticks_t free_time = 0, free_count = 0;
+
+size_t ideal_rss = 0;
+size_t max_ideal_rss = 0;
+static pthread_mutex_t rss_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+void atomic_rss (ssize_t delta)
+{
+  pthread_mutex_lock (&rss_mutex);
+  ideal_rss += delta;
+  if (max_ideal_rss < ideal_rss)
+    max_ideal_rss = ideal_rss;
+  pthread_mutex_unlock (&rss_mutex);
+}
+
+pthread_mutex_t stop_mutex = PTHREAD_MUTEX_INITIALIZER;
+int threads_done = 0;
+
+//#define dprintf printf
+#define dprintf(...) (void)1
+
+//#define mprintf printf
+//#define MDEBUG 1
+#define mprintf(...) (void)1
+
+#define myabort() my_abort_2(thrc, __LINE__)
+void
+my_abort_2 (pthread_t thrc, int line)
+{
+  fprintf(stderr, "Abort thread %p at line %d\n", (void *)thrc, line);
+  abort();
+}
+
+/*------------------------------------------------------------*/
+/* Wrapper around I/O routines */
+
+int io_fd;
+
+#define IOSIZE 65536
+#define IOMIN 4096
+
+static pthread_mutex_t io_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+typedef struct {
+  unsigned char buf[IOSIZE];
+  size_t incr;
+  size_t max_incr;
+  size_t buf_base;
+  size_t buf_idx;
+  int saw_eof;
+} IOPerThreadType;
+
+IOPerThreadType main_io;
+IOPerThreadType *thread_io;
+
+void
+io_init (IOPerThreadType *io, size_t file_offset, int incr)
+{
+  if (incr > IOSIZE)
+    incr = IOSIZE;
+  if (incr < IOMIN)
+    incr = IOMIN;
+
+  io->buf_base = file_offset;
+  io->buf_idx = 0;
+  io->incr = incr;
+
+  pthread_mutex_lock (&io_mutex);
+  lseek64 (io_fd, io->buf_base, SEEK_SET);
+  // short read OK, the eof is just to prevent runaways from bad data.
+  if (read (io_fd, io->buf, incr) < 0)
+    io->saw_eof = 1;
+  else
+    io->saw_eof = 0;
+  pthread_mutex_unlock (&io_mutex);
+}
+
+unsigned char
+io_read (IOPerThreadType *io)
+{
+  if (io->buf_idx >= io->incr)
+    io_init (io, io->buf_base + io->buf_idx, io->incr);
+  if (io->saw_eof)
+    return 0xff;
+  return io->buf [io->buf_idx++];
+}
+
+unsigned char
+io_peek (IOPerThreadType *io)
+{
+  if (io->buf_idx >= io->incr)
+    io_init (io, io->buf_base + io->buf_idx, io->incr);
+  if (io->saw_eof)
+    return 0xff;
+  return io->buf [io->buf_idx];
+}
+
+size_t
+io_pos (IOPerThreadType *io)
+{
+  return io->buf_base + io->buf_idx;
+}
+
+/*------------------------------------------------------------*/
+
+static void
+wmem (volatile void *ptr, int count)
+{
+  char *p = (char *)ptr;
+  int i;
+
+  if (!p)
+    return;
+
+  for (i=0; i<count; i++)
+    p[i] = 0x11;
+}
+#define xwmem(a,b)
+
+static size_t get_int (IOPerThreadType *io)
+{
+  size_t rv = 0;
+  while (1)
+  {
+    unsigned char c = io_read (io);
+    rv |= (c & 0x7f);
+    if (c & 0x80)
+      rv <<= 7;
+    else
+      return rv;
+  }
+}
+
+static void free_wipe (size_t idx)
+{
+  char *cp = (char *)ptrs[idx];
+  if (cp == NULL)
+    return;
+  size_t sz = sizes[idx];
+  size_t i;
+  for (i=0; i<sz; i++)
+    {
+      if (i % 8 == 1)
+	cp[i] = i / 8;
+      else
+	cp[i] = 0x22;
+    }
+}
+
+static void *
+thread_common (void *my_data_v)
+{
+  pthread_t thrc = pthread_self ();
+  size_t p1, p2, sz, sz2;
+  IOPerThreadType *io = (IOPerThreadType *)my_data_v;
+  ticks_t my_malloc_time = 0, my_malloc_count = 0;
+  ticks_t my_calloc_time = 0, my_calloc_count = 0;
+  ticks_t my_realloc_time = 0, my_realloc_count = 0;
+  ticks_t my_free_time = 0, my_free_count = 0;
+  ticks_t stime, etime;
+  int thread_idx = io - thread_io;
+#ifdef MDEBUG
+  volatile void *tmp;
+#endif
+
+  while (1)
+    {
+      unsigned char this_op = io_peek (io);
+      if (io->saw_eof)
+	myabort();
+      dprintf("op %p:%ld is %d\n", (void *)thrc, io_pos (io),  io_peek (io));
+      switch (io_read (io))
+	{
+	case C_NOP:
+	  break;
+
+	case C_DONE:
+	  dprintf("op %p:%ld DONE\n", (void *)thrc, io_pos (io));
+	  pthread_mutex_lock (&stat_mutex);
+	  malloc_time += my_malloc_time;
+	  calloc_time += my_calloc_time;
+	  realloc_time += my_realloc_time;
+	  free_time += my_free_time;
+	  malloc_count += my_malloc_count;
+	  calloc_count += my_calloc_count;
+	  realloc_count += my_realloc_count;
+	  free_count += my_free_count;
+	  threads_done ++;
+	  pthread_mutex_unlock (&stat_mutex);
+	  pthread_mutex_lock(&stop_mutex);
+	  pthread_mutex_unlock(&stop_mutex);
+	  return NULL;
+
+	case C_MEMALIGN:
+	  p2 = get_int (io);
+	  sz2 = get_int (io);
+	  sz = get_int (io);
+	  dprintf("op %p:%ld %ld = MEMALIGN %ld %ld\n", (void *)thrc, io_pos (io), p2, sz2, sz);
+	  /* we can't force memalign to return NULL (fail), so just skip it.  */
+	  if (p2 == 0)
+	    break;
+	  if (p2 > n_ptrs)
+	    myabort();
+	  stime = rdtsc_s();
+	  Q1;
+	  if (ptrs[p2])
+	    {
+	      if (!quick_run)
+		free ((void *)ptrs[p2]);
+	      atomic_rss (-sizes[p2]);
+	    }
+	  if (!quick_run)
+	    ptrs[p2] = memalign (sz2, sz);
+	  else
+	    ptrs[p2] = (void *)p2;
+	  /* Verify the alignment matches what is expected.  */
+	  if (((size_t)ptrs[p2] & (sz2 - 1)) != 0)
+	    myabort ();
+	  sizes[p2] = sz;
+	  mprintf("%p = memalign(%lx, %lx)\n", ptrs[p2], sz2, sz);
+	  Q2;
+	  etime = rdtsc_e();
+	  if (ptrs[p2] != NULL)
+	    atomic_rss (sz);
+	  if (etime < stime)
+	    {
+	      printf("s: %llx e:%llx  d:%llx\n", (long long)stime, (long long)etime, (long long)(etime-stime));
+	    }
+	  my_malloc_time += etime - stime;
+	  my_malloc_count ++;
+	  if (!quick_run)
+	    wmem(ptrs[p2], sz);
+	  break;
+
+	case C_MALLOC:
+	  p2 = get_int (io);
+	  sz = get_int (io);
+	  dprintf("op %p:%ld %ld = MALLOC %ld\n", (void *)thrc, io_pos (io), p2, sz);
+	  /* we can't force malloc to return NULL (fail), so just skip it.  */
+	  if (p2 == 0)
+	    break;
+	  if (p2 > n_ptrs)
+	    myabort();
+	  stime = rdtsc_s();
+	  Q1;
+	  if (ptrs[p2])
+	    {
+	      if (!quick_run)
+		free ((void *)ptrs[p2]);
+	      atomic_rss (-sizes[p2]);
+	    }
+	  if (!quick_run)
+	    ptrs[p2] = malloc (sz);
+	  else
+	    ptrs[p2] = (void *)p2;
+	  sizes[p2] = sz;
+	  mprintf("%p = malloc(%lx)\n", ptrs[p2], sz);
+	  Q2;
+	  etime = rdtsc_e();
+	  if (ptrs[p2] != NULL)
+	    atomic_rss (sz);
+	  if (etime < stime)
+	    {
+	      printf("s: %llx e:%llx  d:%llx\n", (long long)stime, (long long)etime, (long long)(etime-stime));
+	    }
+	  my_malloc_time += etime - stime;
+	  my_malloc_count ++;
+	  if (!quick_run)
+	    wmem(ptrs[p2], sz);
+	  break;
+
+	case C_CALLOC:
+	  p2 = get_int (io);
+	  sz = get_int (io);
+	  dprintf("op %p:%ld %ld = CALLOC %ld\n", (void *)thrc, io_pos (io), p2, sz);
+	  /* we can't force calloc to return NULL (fail), so just skip it.  */
+	  if (p2 == 0)
+	    break;
+	  if (p2 > n_ptrs)
+	    myabort();
+	  if (ptrs[p2])
+	    {
+	      if (!quick_run)
+		free ((void *)ptrs[p2]);
+	      atomic_rss (-sizes[p2]);
+	    }
+	  stime = rdtsc_s();
+	  Q1;
+	  if (!quick_run)
+	    ptrs[p2] = calloc (sz, 1);
+	  else
+	    ptrs[p2] = (void *)p2;
+	  sizes[p2] = sz;
+	  mprintf("%p = calloc(%lx)\n", ptrs[p2], sz);
+	  Q2;
+	  if (ptrs[p2])
+	    atomic_rss (sz);
+	  my_calloc_time += rdtsc_e() - stime;
+	  my_calloc_count ++;
+	  if (!quick_run)
+	    wmem(ptrs[p2], sz);
+	  break;
+
+	case C_REALLOC:
+	  p2 = get_int (io);
+	  p1 = get_int (io);
+	  sz = get_int (io);
+	  dprintf("op %p:%ld %ld = REALLOC %ld %ld\n", (void *)thrc, io_pos (io), p2, p1, sz);
+	  if (p1 > n_ptrs)
+	    myabort();
+	  if (p2 > n_ptrs)
+	    myabort();
+	  /* we can't force realloc to return NULL (fail), so just skip it.  */
+	  if (p2 == 0)
+	    break;
+
+	  if (ptrs[p1])
+	    atomic_rss (-sizes[p1]);
+	  if (!quick_run)
+	    free_wipe(p1);
+	  stime = rdtsc_s();
+	  Q1;
+#ifdef MDEBUG
+	  tmp = ptrs[p1];
+#endif
+	  if (!quick_run)
+	    ptrs[p2] = realloc ((void *)ptrs[p1], sz);
+	  else
+	    ptrs[p2] = (void *)p2;
+	  sizes[p2] = sz;
+	  mprintf("%p = relloc(%p,%lx)\n", ptrs[p2], tmp,sz);
+	  Q2;
+	  my_realloc_time += rdtsc_e() - stime;
+	  my_realloc_count ++;
+	  if (!quick_run)
+	    wmem(ptrs[p2], sz);
+	  if (p1 != p2)
+	    ptrs[p1] = 0;
+	  if (ptrs[p2])
+	    atomic_rss (sizes[p2]);
+	  break;
+
+	case C_FREE:
+	  p1 = get_int (io);
+	  if (p1 > n_ptrs)
+	    myabort();
+	  dprintf("op %p:%ld FREE %ld\n", (void *)thrc, io_pos (io), p1);
+	  if (!quick_run)
+	    free_wipe (p1);
+	  if (ptrs[p1])
+	    atomic_rss (-sizes[p1]);
+	  stime = rdtsc_s();
+	  Q1;
+	  mprintf("free(%p)\n", ptrs[p1]);
+	  if (!quick_run)
+	    free ((void *)ptrs[p1]);
+	  Q2;
+	  my_free_time += rdtsc_e() - stime;
+	  my_free_count ++;
+	  ptrs[p1] = 0;
+	  break;
+
+	case C_SYNC_W:
+	  p1 = get_int(io);
+	  dprintf("op %p:%ld SYNC_W %ld\n", (void *)thrc, io_pos (io), p1);
+	  if (p1 > n_syncs)
+	    myabort();
+	  pthread_mutex_lock (&mutexes[p1]);
+	  syncs[p1] = 1;
+	  pthread_cond_signal (&conds[p1]);
+	  __sync_synchronize ();
+	  pthread_mutex_unlock (&mutexes[p1]);
+	  break;
+
+	case C_SYNC_R:
+	  p1 = get_int(io);
+	  dprintf("op %p:%ld SYNC_R %ld\n", (void *)thrc, io_pos (io), p1);
+	  if (p1 > n_syncs)
+	    myabort();
+	  pthread_mutex_lock (&mutexes[p1]);
+	  while (syncs[p1] != 1)
+	    {
+	      pthread_cond_wait (&conds[p1], &mutexes[p1]);
+	      __sync_synchronize ();
+	    }
+	  pthread_mutex_unlock (&mutexes[p1]);
+	  break;
+
+	default:
+	  printf("op %d - unsupported, thread %d addr %lu\n",
+		 this_op, thread_idx, (long unsigned int)io_pos (io));
+	  myabort();
+	}
+    }
+}
+
+static void *alloc_mem (size_t amt)
+{
+  void *rv = mmap (NULL, amt, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+  mlock (rv, amt);
+  memset (rv, 0, amt);
+  return rv;
+}
+
+static pthread_t *thread_ids;
+
+void *
+my_malloc (const char *msg, int size, IOPerThreadType *io, size_t *psz, size_t count)
+{
+  void *rv;
+  if (psz)
+    count = *psz = get_int (io);
+  dprintf ("my_malloc for %s size %d * %ld\n", msg, size, count);
+  rv = alloc_mem(size * count);
+  if (!rv)
+    {
+      fprintf(stderr, "calloc(%lu,%lu) failed\n", (long unsigned)size, (long unsigned)*psz);
+      exit(1);
+    }
+  mlock (rv, size * count);
+  return rv;
+}
+
+static const char * const scan_names[] = {
+  "UNUSED",
+  "ARENA",
+  "HEAP",
+  "CHUNK_USED",
+  "CHUNK_FREE",
+  "FASTBIN_FREE",
+  "UNSORTED",
+  "TOP",
+  "TCACHE",
+  "USED"
+};
+
+void
+malloc_scan_callback (void *ptr, size_t length, int type)
+{
+  printf("%s: ptr %p length %llx\n", scan_names[type], ptr, (long long)length);
+}
+
+#define MY_ALLOC(T, psz)				\
+  (typeof (T)) my_malloc (#T, sizeof(*T), &main_io, psz, 0)
+#define MY_ALLOCN(T, count)				\
+  (typeof (T)) my_malloc (#T, sizeof(*T), &main_io, NULL, count)
+
+int
+main(int argc, char **argv)
+{
+  ticks_t start=0;
+  ticks_t end;
+  ticks_t usec;
+  struct timeval tv_s, tv_e;
+  int thread_idx = 0;
+  int i;
+  size_t n_threads = 0;
+  size_t idx;
+  struct rusage res_start, res_end;
+  int done;
+  size_t guessed_io_size = 4096;
+  struct stat statb;
+
+  if (argc < 2)
+    {
+      fprintf(stderr, "Usage: %s <trace2dat.outfile>\n", argv[0]);
+      exit(1);
+    }
+  io_fd = open(argv[1], O_RDONLY);
+  if (io_fd < 0)
+    {
+      fprintf(stderr, "Unable to open %s for reading\n", argv[1]);
+      perror("The error was");
+      exit(1);
+    }
+  fstat (io_fd, &statb);
+
+  io_init (&main_io, 0, IOMIN);
+
+  pthread_mutex_lock(&stop_mutex);
+
+  done = 0;
+  while (!done)
+    {
+      switch (io_read (&main_io))
+	{
+	case C_NOP:
+	  break;
+	case C_ALLOC_PTRS:
+	  ptrs = MY_ALLOC (ptrs, &n_ptrs);
+	  sizes = alloc_mem(sizeof(sizes[0]) * n_ptrs);
+	  ptrs[0] = 0;
+	  break;
+	case C_ALLOC_SYNCS:
+	  n_syncs = get_int(&main_io);
+	  syncs = MY_ALLOCN (syncs, n_syncs);
+	  conds = MY_ALLOCN (conds, n_syncs);
+	  mutexes = MY_ALLOCN (mutexes, n_syncs);
+	  for (idx=0; idx<n_syncs; idx++)
+	    {
+	      pthread_mutex_init (&mutexes[idx], NULL);
+	      pthread_cond_init (&conds[idx], NULL);
+	    }
+	  break;
+	case C_NTHREADS:
+	  thread_ids = MY_ALLOC (thread_ids, &n_threads);
+	  thread_io = MY_ALLOCN (thread_io, n_threads);
+	  guessed_io_size = ((statb.st_size / n_threads) < (1024*1024)) ? 65536 : 4096;
+	  /* The next thing in the workscript is thread creation */
+	  getrusage (RUSAGE_SELF, &res_start);
+	  gettimeofday (&tv_s, NULL);
+	  start = rdtsc_s();
+	  break;
+	case C_START_THREAD:
+	  idx = get_int (&main_io);
+	  io_init (& thread_io[thread_idx], idx, guessed_io_size);
+	  pthread_create (&thread_ids[thread_idx], NULL, thread_common, thread_io + thread_idx);
+	  dprintf("Starting thread %lld at offset %lu %lx\n", (long long)thread_ids[thread_idx], (unsigned long)idx, (unsigned long)idx);
+	  thread_idx ++;
+	  break;
+	case C_DONE:
+	  do
+	    {
+	      pthread_mutex_lock (&stat_mutex);
+	      i = threads_done;
+	      pthread_mutex_unlock (&stat_mutex);
+	    } while (i < thread_idx);
+	  done = 1;
+	  break;
+	}
+    }
+  if (!quick_run)
+    {
+      end = rdtsc_e();
+      gettimeofday (&tv_e, NULL);
+      getrusage (RUSAGE_SELF, &res_end);
+
+      printf("%s cycles\n", comma(end - start));
+      usec = diff_timeval (tv_e, tv_s);
+      printf("%s usec wall time\n", comma(usec));
+
+      usec = diff_timeval (res_end.ru_utime, res_start.ru_utime);
+      printf("%s usec across %d thread%s\n",
+	     comma(usec), (int)n_threads, n_threads == 1 ? "" : "s");
+      printf("%s Kb Max RSS (%s -> %s)\n",
+	     comma(res_end.ru_maxrss - res_start.ru_maxrss),
+	     comma(res_start.ru_maxrss), comma(res_end.ru_maxrss));
+    }
+  printf("%s Kb Max Ideal RSS\n", comma (max_ideal_rss / 1024));
+
+  if (malloc_count == 0) malloc_count ++;
+  if (calloc_count == 0) calloc_count ++;
+  if (realloc_count == 0) realloc_count ++;
+  if (free_count == 0) free_count ++;
+
+  if (!quick_run)
+    {
+      printf("\n");
+      printf("sizeof ticks_t is %lu\n", sizeof(ticks_t));
+      printf("Avg malloc time: %6s in %10s calls\n", comma(malloc_time/malloc_count), comma(malloc_count));
+      printf("Avg calloc time: %6s in %10s calls\n", comma(calloc_time/calloc_count), comma(calloc_count));
+      printf("Avg realloc time: %5s in %10s calls\n", comma(realloc_time/realloc_count), comma(realloc_count));
+      printf("Avg free time: %8s in %10s calls\n", comma(free_time/free_count), comma(free_count));
+      printf("Total call time: %s cycles\n", comma(malloc_time+calloc_time+realloc_time+free_time));
+      printf("\n");
+    }
+
+#if 0
+  /* Free any still-held chunks of memory.  */
+  for (idx=0; idx<n_ptrs; idx++)
+    if (ptrs[idx])
+      {
+	free((void *)ptrs[idx]);
+	ptrs[idx] = 0;
+      }
+#endif
+
+#if 0
+  /* This will fail (crash) for system glibc but that's OK.  */
+  __malloc_scan_chunks(malloc_scan_callback);
+
+  malloc_info (0, stdout);
+#endif
+
+#if 0
+  /* ...or report them as used.  */
+  for (idx=0; idx<n_ptrs; idx++)
+    if (ptrs[idx])
+      {
+	char *p = (char *)ptrs[idx] - 2*sizeof(size_t);
+	size_t *sp = (size_t *)p;
+	size_t size = sp[1] & ~7;
+	malloc_scan_callback (sp, size, 9);
+      }
+#endif
+
+  /* Now that we've scanned all the per-thread caches, it's safe to
+     let them exit and clean up.  */
+  pthread_mutex_unlock(&stop_mutex);
+
+  for (i=0; i<thread_idx; i++)
+    pthread_join (thread_ids[i], NULL);
+
+  return 0;
+}