Measure the difference between predicted execution time and measured execution

time.

src/core/mechanisms/queues.h

@@ -52,6 +52,7 @@ struct jobq_s {
 	/* for performance analysis purpose */
 	double total_computation_time;
 	double total_communication_time;
+	double total_computation_time_error;
 
 	/* in case workers are blocked on the queue, signaling on that 
 	  condition must unblock them, even if there is no available task */

src/drivers/core/driver_core.c

@@ -14,6 +14,8 @@
  * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  */
 
+#include <math.h>
+
 #include "driver_core.h"
 #include <core/policies/sched_policy.h>
 
@@ -72,6 +74,11 @@ int execute_job_on_core(job_t j, struct worker_s *core_args)
 		core_args->jobq->total_computation_time += measured;
 		core_args->jobq->total_communication_time += measured_comm;
 
+		double error;
+		error = fabs(STARPU_MAX(measured, 0.0) - STARPU_MAX(j->predicted, 0.0)); 
+//		fprintf(stderr, "Error -> %le, predicted -> %le measured ->%le\n", error, j->predicted, measured);
+		core_args->jobq->total_computation_time_error += error;
+
 		if (calibrate_model)
 			update_perfmodel_history(j, core_args->arch, core_args->id, measured);
 	}
@@ -105,6 +112,7 @@ void *core_worker(void *arg)
 
 	core_arg->jobq->total_computation_time = 0.0;
 	core_arg->jobq->total_communication_time = 0.0;
+	core_arg->jobq->total_computation_time_error = 0.0;
 	
         /* tell the main thread that we are ready */
 	pthread_mutex_lock(&core_arg->mutex);
@@ -148,6 +156,10 @@ void *core_worker(void *arg)
 	fprintf(stderr, "CORE #%d computation %le comm %le (%lf \%%)\n", core_arg->id, core_arg->jobq->total_computation_time, core_arg->jobq->total_communication_time,  core_arg->jobq->total_communication_time*100.0/core_arg->jobq->total_computation_time);
 #endif
 
+#ifdef VERBOSE
+	fprintf(stderr, "CORE #%d error %le error/exec %le\n", core_arg->id, core_arg->jobq->total_computation_time_error,  core_arg->jobq->total_computation_time_error/core_arg->jobq->total_computation_time);
+#endif
+
 	TRACE_WORKER_TERMINATED(FUT_CORE_KEY);
 
 	pthread_exit(NULL);

src/drivers/cuda/driver_cuda.c

@@ -307,6 +307,10 @@ int execute_job_on_cuda(job_t j, struct worker_s *args, unsigned use_cublas)
 		args->jobq->total_computation_time += measured;
 		args->jobq->total_communication_time += measured_comm;
 
+		double error;
+		error = fabs(STARPU_MAX(measured, 0.0) - STARPU_MAX(j->predicted, 0.0)); 
+		args->jobq->total_computation_time_error += error;
+
 		if (calibrate_model)
 			update_perfmodel_history(j, args->arch, args->id, measured);
 	}
@@ -341,6 +345,7 @@ void *cuda_worker(void *arg)
 
 	args->jobq->total_computation_time = 0.0;
 	args->jobq->total_communication_time = 0.0;
+	args->jobq->total_computation_time_error = 0.0;
 
 	init_context(devid);
 #ifdef VERBOSE
@@ -412,6 +417,10 @@ void *cuda_worker(void *arg)
 #endif
 	pthread_exit(NULL);
 
+#ifdef VERBOSE
+	fprintf(stderr, "CORE #%d error %le error/exec %le\n", args->id, args->jobq->total_computation_time_error, args->jobq->total_computation_time_error/args->jobq->total_computation_time );
+#endif
+
 	TRACE_WORKER_TERMINATED(FUT_CUDA_KEY);
 
 	return NULL;