diff --git a/Makefile b/Makefile
index a5f350e..c16c206 100644
--- a/Makefile
+++ b/Makefile
@@ -3,7 +3,7 @@ CFLAGS=-g -O0 -fPIC
 CFLAGS+=-fsanitize=address
 LDFLAGS+=-fsanitize=address
 
-BIN=dump libovni.a
+BIN=dump libovni.a prvth
 
 all: $(BIN)
 
diff --git a/def.h b/def.h
index ffab9f6..701924d 100644
--- a/def.h
+++ b/def.h
@@ -18,7 +18,8 @@ struct __attribute__((__packed__)) event {
 	uint64_t clock;
 	uint8_t fsm;
 	uint8_t event;
-	int32_t data;
+	uint16_t a;
+	uint16_t b;
 };
 
 /* ----------------------- runtime ------------------------ */
@@ -38,7 +39,7 @@ struct rthread {
 	/* Stream trace file */
 	FILE *stream;
 
-	enum thread_state state;
+	int ready;
 };
 
 /* State of each process on runtime */
@@ -51,6 +52,8 @@ struct rproc {
 	int ncpus;
 	clockid_t clockid;
 	char procdir[PATH_MAX];
+
+	int ready;
 };
 
 /* ----------------------- emulated ------------------------ */
@@ -91,10 +94,14 @@ struct loom {
 
 struct stream {
 	FILE *f;
-	int cpu;
+	int tid;
+	int thread;
+	int proc;
+	int loom;
 	int loaded;
 	int active;
 	struct event last;
+	uint64_t lastclock;
 };
 
 struct trace {
diff --git a/dump.c b/dump.c
index c48f76d..d23f5c9 100644
--- a/dump.c
+++ b/dump.c
@@ -121,6 +121,7 @@ int load_streams(struct trace *trace)
 	int i, j, k, s;
 	struct loom *loom;
 	struct eproc *proc;
+	struct ethread *thread;
 	struct stream *stream;
 
 	trace->nstreams = 0;
@@ -156,9 +157,15 @@ int load_streams(struct trace *trace)
 			proc = &loom->proc[j];
 			for(k=0; k<proc->nthreads; k++)
 			{
+				thread = &proc->thread[k];
 				stream = &trace->stream[s++];
-				stream->f = proc->thread[k].f;
+
+				stream->f = thread->f;
+				stream->tid = thread->tid;
+				stream->proc = j;
+				stream->loom = i;
 				stream->active = 1;
+				stream->lastclock = 0;
 			}
 		}
 	}
@@ -188,17 +195,17 @@ int load_first_event(struct trace *trace)
 	return 0;
 }
 
-void emit(struct event *ev, int cpu)
+void emit(struct stream *stream, struct event *ev)
 {
-	static uint64_t lastclock = 0;
-	uint64_t delta;
+	int64_t delta;
 
-	delta = ev->clock - lastclock;
+	delta = ev->clock - stream->lastclock;
 
-	printf("%d %c %c %d %020lu (+%lu) ns\n",
-			cpu, ev->fsm, ev->event, ev->data, ev->clock, delta);
+	printf("%d.%d.%d %c %c % 6u % 6u % 20lu %+15ld ns\n",
+			stream->loom, stream->proc, stream->tid,
+			ev->fsm, ev->event, ev->a, ev->b, ev->clock, delta);
 
-	lastclock = ev->clock;
+	stream->lastclock = ev->clock;
 }
 
 void load_next_event(struct stream *stream)
@@ -271,7 +278,7 @@ void dump_events(struct trace *trace)
 		}
 
 		/* Emit current event */
-		emit(&stream->last, stream->cpu);
+		emit(stream, &stream->last);
 
 		lastclock = stream->last.clock;
 
diff --git a/ovni.c b/ovni.c
index 8909e41..2063339 100644
--- a/ovni.c
+++ b/ovni.c
@@ -1,3 +1,6 @@
+#define _GNU_SOURCE
+#include <unistd.h>
+
 #include <stdio.h>
 #include <stdint.h>
 #include <stdlib.h>
@@ -58,10 +61,14 @@ create_trace_stream()
 {
 	char path[PATH_MAX];
 
+	fprintf(stderr, "create thread stream tid=%d gettid=%d rproc.proc=%d rproc.ready=%d\n",
+			rthread.tid, gettid(), rproc.proc, rproc.ready);
+
 	snprintf(path, PATH_MAX, "%s/thread.%d", rproc.dir, rthread.tid);
 	if((rthread.stream = fopen(path, "w")) == NULL)
 	{
 		fprintf(stderr, "fopen %s failed: %s\n", path, strerror(errno));
+		abort();
 		return -1;
 	}
 
@@ -69,20 +76,24 @@ create_trace_stream()
 }
 
 int
-ovni_proc_init(int loom, int proc, int ncpus)
+ovni_proc_init(int loom, int proc)
 {
 	int i;
+
+	assert(rproc.ready == 0);
+
 	memset(&rproc, 0, sizeof(rproc));
 
 	rproc.loom = loom;
 	rproc.proc = proc;
-	rproc.ncpus = ncpus;
 
 	/* By default we use the monotonic clock */
 	rproc.clockid = CLOCK_MONOTONIC;
 
 	if(create_trace_dirs(TRACEDIR, loom, proc))
-		return -1;
+		abort();
+
+	rproc.ready = 1;
 
 	return 0;
 }
@@ -94,18 +105,39 @@ ovni_thread_init(pid_t tid)
 
 	assert(tid != 0);
 
+	if(rthread.ready)
+	{
+		fprintf(stderr, "warning: thread tid=%d already initialized\n",
+				tid);
+		return 0;
+	}
+
+	assert(rthread.ready == 0);
+	assert(rthread.tid == 0);
+	assert(rthread.cpu == 0);
+	assert(rproc.ready == 1);
+
+	fprintf(stderr, "ovni thread init tid=%d\n", tid);
+
 	memset(&rthread, 0, sizeof(rthread));
 
 	rthread.tid = tid;
-	rthread.cpu = -1;
-	rthread.state = ST_THREAD_INIT;
+	rthread.cpu = -666;
 
 	if(create_trace_stream(tid))
-		return -1;
+		abort();
+
+	rthread.ready = 1;
 
 	return 0;
 }
 
+int
+ovni_thread_isready()
+{
+	return rthread.ready;
+}
+
 void
 ovni_cpu_set(int cpu)
 {
@@ -151,6 +183,8 @@ hexdump(uint8_t *buf, size_t size)
 static int
 ovni_write(uint8_t *buf, size_t size)
 {
+	assert(rthread.stream);
+
 	fprintf(stderr, "writing %ld bytes in thread.%d\n", size, rthread.tid);
 	if(fwrite(buf, 1, size, rthread.stream) != size)
 	{
@@ -158,27 +192,29 @@ ovni_write(uint8_t *buf, size_t size)
 		return -1;
 	}
 
-	fflush(rthread.stream);
+	//fflush(rthread.stream);
 
 	return 0;
 }
 
 static int
-ovni_ev(uint8_t fsm, uint8_t event, int32_t data)
+ovni_ev(uint8_t fsm, uint8_t event, uint16_t a, uint16_t b)
 {
 	struct event ev;
 
 	ev.clock = rthread.clockvalue;
 	ev.fsm = fsm;
 	ev.event = event;
-	ev.data = data;
+	ev.a = a;
+	ev.b = b;
 
 	return ovni_write((uint8_t *) &ev, sizeof(ev));
 }
 
 int
-ovni_ev_worker(uint8_t fsm, uint8_t event, int32_t data)
+ovni_thread_ev(uint8_t fsm, uint8_t event, uint16_t a, uint16_t b)
 {
-	assert(rthread.state == ST_THREAD_INIT);
-	return ovni_ev(fsm, event, data);
+	assert(rthread.ready);
+	assert(rproc.ready);
+	return ovni_ev(fsm, event, a, b);
 }
diff --git a/ovni.h b/ovni.h
index 677fc3d..33a2171 100644
--- a/ovni.h
+++ b/ovni.h
@@ -2,18 +2,18 @@
 #define OVNI_H
 
 int
-ovni_proc_init(int loom, int proc, int ncpus);
+ovni_proc_init(int loom, int proc);
 
 int
 ovni_thread_init(pid_t tid);
 
+int
+ovni_thread_isready();
+
 int
 ovni_clock_update();
 
 int
-ovni_ev_worker(uint8_t fsm, uint8_t event, int32_t data);
-
-void
-ovni_cpu_set(int cpu);
+ovni_thread_ev(uint8_t fsm, uint8_t event, uint16_t a, uint16_t b);
 
 #endif /* OVNI_H */
diff --git a/prvth.c b/prvth.c
new file mode 100644
index 0000000..d5987fd
--- /dev/null
+++ b/prvth.c
@@ -0,0 +1,336 @@
+#include <stdio.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <linux/limits.h>
+#include <errno.h>
+#include <sys/stat.h>
+#include <stdatomic.h>
+#include <dirent.h> 
+
+#include "def.h"
+
+int load_proc(struct eproc *proc, char *procdir)
+{
+	struct dirent *dirent;
+	DIR *dir;
+	char path[PATH_MAX];
+	char *p;
+	struct ethread *thread;
+
+	if((dir = opendir(procdir)) == NULL)
+	{
+		fprintf(stderr, "opendir %s failed: %s\n",
+				procdir, strerror(errno));
+		return -1;
+	}
+
+	while((dirent = readdir(dir)) != NULL)
+	{
+		if(dirent->d_name[0] != 't')
+			continue;
+		p = strchr(dirent->d_name, '.');
+		if(p == NULL)
+			continue;
+		p++;
+		if(*p == '\0')
+		{
+			fprintf(stderr, "bad thread stream file: %s\n",
+					dirent->d_name);
+			return -1;
+		}
+
+
+		thread = &proc->thread[proc->nthreads++];
+		thread->tid = atoi(p);
+
+		sprintf(path, "%s/%s", procdir, dirent->d_name);
+		thread->f = fopen(path, "r");
+
+		if(thread->f == NULL)
+		{
+			fprintf(stderr, "fopen %s failed: %s\n",
+					path, strerror(errno));
+			return -1;
+		}
+	}
+
+	closedir(dir);
+
+	return 0;
+}
+
+int load_loom(struct loom *loom, char *loomdir)
+{
+	int proc;
+	char path[PATH_MAX];
+	struct stat st;
+
+	for(proc=0; proc<MAX_PROC; proc++)
+	{
+		sprintf(path, "%s/proc.%d", loomdir, proc);
+
+		if(stat(path, &st) != 0)
+		{
+			/* No more proc.N directories */
+			if(errno == ENOENT)
+				break;
+
+			fprintf(stderr, "cannot stat %s: %s\n", path,
+					strerror(errno));
+			return -1;
+		}
+
+		if(!S_ISDIR(st.st_mode))
+		{
+			fprintf(stderr, "not a dir %s\n", path);
+			return -1;
+		}
+
+		if(load_proc(&loom->proc[proc], path))
+			return -1;
+	}
+
+	loom->nprocs = proc;
+
+	return 0;
+}
+
+int load_trace(struct trace *trace, char *tracedir)
+{
+	int loom, nlooms;
+	char path[PATH_MAX];
+
+	/* TODO: For now only one loom */
+	nlooms = 1;
+	loom = 0;
+
+	sprintf(path, "%s/loom.%d", tracedir, loom);
+
+	if(load_loom(&trace->loom[loom], path))
+		return -1;
+
+	trace->nlooms = nlooms;
+
+	return 0;
+}
+
+/* Populates the streams in a single array */
+int load_streams(struct trace *trace)
+{
+	int i, j, k, s;
+	struct loom *loom;
+	struct eproc *proc;
+	struct ethread *thread;
+	struct stream *stream;
+
+	trace->nstreams = 0;
+
+	for(i=0; i<trace->nlooms; i++)
+	{
+		loom = &trace->loom[i];
+		for(j=0; j<loom->nprocs; j++)
+		{
+			proc = &loom->proc[j];
+			for(k=0; k<proc->nthreads; k++)
+			{
+				trace->nstreams++;
+			}
+		}
+	}
+
+	trace->stream = calloc(trace->nstreams, sizeof(struct stream));
+
+	if(trace->stream == NULL)
+	{
+		perror("calloc");
+		return -1;
+	}
+
+	fprintf(stderr, "loaded %d streams\n", trace->nstreams);
+
+	for(s=0, i=0; i<trace->nlooms; i++)
+	{
+		loom = &trace->loom[i];
+		for(j=0; j<loom->nprocs; j++)
+		{
+			proc = &loom->proc[j];
+			for(k=0; k<proc->nthreads; k++)
+			{
+				thread = &proc->thread[k];
+				stream = &trace->stream[s++];
+
+				stream->f = thread->f;
+				stream->tid = thread->tid;
+				stream->thread = k;
+				stream->proc = j;
+				stream->loom = i;
+				stream->active = 1;
+			}
+		}
+	}
+
+	return 0;
+}
+
+int load_first_event(struct trace *trace)
+{
+	int i;
+	struct stream *stream;
+	struct event *ev;
+
+	for(i=0; i<trace->nstreams; i++)
+	{
+		stream = &trace->stream[i];
+		ev = &stream->last;
+		if(fread(ev, sizeof(*ev), 1, stream->f) != 1)
+		{
+			fprintf(stderr, "failed to read and event\n");
+			return -1;
+		}
+
+		fprintf(stderr, "ev clock %u\n", ev->clock);
+	}
+
+	return 0;
+}
+
+void emit(struct stream *stream, struct event *ev)
+{
+	static uint64_t firstclock = 0;
+	int64_t delta;
+	int task;
+
+	if(firstclock == 0)
+		firstclock = ev->clock;
+
+	delta = ev->clock - firstclock;
+
+	//#Paraver (19/01/38 at 03:14):00000000000000000000_ns:0:1:1(00000000000000000008:1)
+	//2:0:1:1:7:1540663:6400010:1
+	//2:0:1:1:7:1540663:6400015:1
+	//2:0:1:1:7:1540663:6400017:0
+	//2:0:1:1:7:1542091:6400010:1
+	//2:0:1:1:7:1542091:6400015:1
+	//2:0:1:1:7:1542091:6400025:1
+	//2:0:1:1:7:1542091:6400017:0
+
+	printf("2:0:1:1:%d:%ld:%d:%d\n", stream->thread+1, delta, ev->fsm, ev->event);
+}
+
+void load_next_event(struct stream *stream)
+{
+	int i;
+	size_t n;
+	struct event *ev;
+
+	if(!stream->active)
+		return;
+
+	ev = &stream->last;
+	if((n = fread(ev, sizeof(*ev), 1, stream->f)) != 1)
+	{
+		//fprintf(stderr, "failed to read an event, n=%ld\n", n);
+		stream->active = 0;
+		return;
+	}
+
+	stream->active = 1;
+}
+
+void dump_events(struct trace *trace)
+{
+	int i, f;
+	uint64_t minclock, lastclock;
+	struct event *ev;
+	struct stream *stream;
+
+	/* Load events */
+	for(i=0; i<trace->nstreams; i++)
+	{
+		stream = &trace->stream[i];
+		load_next_event(stream);
+	}
+
+	lastclock = 0;
+
+	while(1)
+	{
+		f = -1;
+		minclock = 0;
+
+		/* Select next event based on the clock */
+		for(i=0; i<trace->nstreams; i++)
+		{
+			stream = &trace->stream[i];
+
+			if(!stream->active)
+				continue;
+
+			ev = &stream->last;
+			if(f < 0 || ev->clock < minclock)
+			{
+				f = i;
+				minclock = ev->clock;
+			}
+		}
+
+		//fprintf(stderr, "f=%d minclock=%u\n", f, minclock);
+
+		if(f < 0)
+			break;
+
+		stream = &trace->stream[f];
+
+		if(lastclock >= stream->last.clock)
+		{
+			fprintf(stderr, "warning: backwards jump in time\n");
+		}
+
+		/* Emit current event */
+		emit(stream, &stream->last);
+
+		lastclock = stream->last.clock;
+
+		/* Read the next one */
+		load_next_event(stream);
+
+		/* Unset the index */
+		f = -1;
+		minclock = 0;
+
+	}
+}
+
+void free_streams(struct trace *trace)
+{
+	free(trace->stream);
+}
+
+int main(int argc, char *argv[])
+{
+	char *tracedir;
+	struct trace trace;
+
+	if(argc != 2)
+	{
+		fprintf(stderr, "missing tracedir\n");
+		exit(EXIT_FAILURE);
+	}
+
+	tracedir = argv[1];
+
+	if(load_trace(&trace, tracedir))
+		return 1;
+
+	if(load_streams(&trace))
+		return 1;
+
+	printf("#Paraver (19/01/38 at 03:14):00000000000000000000_ns:0:1:1(%d:1)\n", trace.nstreams);
+
+	dump_events(&trace);
+
+	free_streams(&trace);
+
+	return 0;
+}