[Starpu-devel] StarPU 1.3 RC1 + peager problem

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

I have been experiencing some problems with the new StarPU 1.3.0 release 
candidate and the peager scheduler. I have reported similar/related 
problems earlier but I am quite sure that this problem did not exists 
with the older StarPU 1.3 nightly snapshots. StarPU 1.2.6 works just 
fine.

The problem can reproduced with the following code:

////////////////////////////////////////////////////////////////////////////////

#include <unistd.h>
#include <unistd.h>
#include <starpu.h>

static void kernel(void *buffers[], void *cl_args)
{
    sleep(1);
}

static struct starpu_codelet codelet = {
    .cpu_funcs = { kernel },
    .type = STARPU_FORKJOIN,
    .max_parallelism = INT_MAX
};

int main()
{
    struct starpu_conf conf;
    starpu_conf_init(&conf);
    conf.ncuda = 0;

    if (starpu_init(&conf) != 0)
        return 1;

    int worker_ids[STARPU_NMAXWORKERS];
    int worker_count = starpu_worker_get_ids_by_type(
        STARPU_CPU_WORKER, worker_ids, STARPU_NMAXWORKERS);

    int parallel_ctx_size = worker_count/2;
    int other_ctx_size = worker_count - parallel_ctx_size;

    unsigned parallel_ctx = starpu_sched_ctx_create(
        worker_ids, parallel_ctx_size, "parallel_cxt",
        STARPU_SCHED_CTX_POLICY_NAME, "peager", 0);

    unsigned other_ctx = starpu_sched_ctx_create(
        worker_ids+parallel_ctx_size, other_ctx_size, "other_ctx",
        STARPU_SCHED_CTX_POLICY_NAME, "eager", 0);

    for (int i = 0; i < 10; i++)
        starpu_task_insert(&codelet, STARPU_SCHED_CTX, parallel_ctx, 0);

    for (int i = 0; i < 10; i++)
        starpu_task_insert(&codelet, STARPU_SCHED_CTX, other_ctx, 0);

    starpu_task_wait_for_all_in_ctx(parallel_ctx);
    starpu_task_wait_for_all_in_ctx(other_ctx);

    starpu_sched_ctx_delete(parallel_ctx);
    starpu_sched_ctx_delete(other_ctx);

    starpu_shutdown();

    return 0;
}

////////////////////////////////////////////////////////////////////////////////

GDB outputs the following:

////////////////////////////////////////////////////////////////////////////////

Thread 8 "CPU 1" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7fffcce69700 (LWP 29747)]
0x00007ffff7927082 in pop_task_peager_policy (sched_ctx_id=<optimized 
out>) at sched_policies/parallel_eager.c:280
280     sched_policies/parallel_eager.c: No such file or directory.
(gdb) bt
#0  0x00007ffff7927082 in pop_task_peager_policy 
(sched_ctx_id=<optimized out>) at sched_policies/parallel_eager.c:280
#1  0x00007ffff7911596 in _starpu_pop_task 
(worker=worker@entry=0x7ffff7ba4790 <_starpu_config+3120>) at 
core/sched_policy.c:880
#2  0x00007ffff792cb1f in _starpu_get_worker_task 
(worker=worker@entry=0x7ffff7ba4790 <_starpu_config+3120>, 
workerid=workerid@entry=1, memnode=memnode@entry=0)
    at drivers/driver_common/driver_common.c:377
#3  0x00007ffff7968926 in _starpu_cpu_driver_run_once 
(cpu_worker=cpu_worker@entry=0x7ffff7ba4790 <_starpu_config+3120>) at 
drivers/cpu/driver_cpu.c:304
#4  0x00007ffff7968b4d in _starpu_cpu_worker (arg=0x7ffff7ba4790 
<_starpu_config+3120>) at drivers/cpu/driver_cpu.c:408
#5  0x00007ffff70a36db in start_thread (arg=0x7fffcce69700) at 
pthread_create.c:463
#6  0x00007ffff75e088f in clone () at 
../sysdeps/unix/sysv/linux/x86_64/clone.S:95
(gdb) frame 0
#0  0x00007ffff7927082 in pop_task_peager_policy 
(sched_ctx_id=<optimized out>) at sched_policies/parallel_eager.c:280
280     in sched_policies/parallel_eager.c
(gdb) print common_data
$1 = (struct _starpu_peager_common_data *) 0x0

////////////////////////////////////////////////////////////////////////////////

Best Regards,
Mirko Myllykoski