Provide mechanism to discard data #695
Conversation
devreal
force-pushed
the
parsec-data-discard
branch
from
08ecd2e
to
bbd1448
Compare
Add a function `parsec_data_discard` that releases the data such that the host copy remains intact but does not prevent destruction of the data once all device copies have been released. This keeps the host copy available for device copies to inspect and avoids potential race conditions in the release process. During an eviction, copies of data with a discarded host copy are not transfered but put directly into the lru. Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>
devreal
force-pushed
the
parsec-data-discard
branch
from
bbd1448
to
1dd2d54
Compare
Otherwise we cannot destroy empty or discarded data. Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>
Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>
Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>
parsec/data.c
Outdated
| /* second, mark the host copy as discarded */ | ||
| parsec_data_copy_t *cpu_copy = data->device_copies[0]; | ||
| if (NULL != cpu_copy) { | ||
| cpu_copy->flags = PARSEC_DATA_FLAG_DISCARDED; |
There was a problem hiding this comment.
Three issues here:
- the flag should be set with
|= - you should only release the
data_tis thecpu_copyis valid. You do a test against NULL, you should use it. - this function is not thread-safe. This could not be an issue here, but I see it used by the GPU code without protection either, opening the code to race-conditions while manipulating these discarded copies/
There was a problem hiding this comment.
Good catch. I fixed setting the flag, added a lock to protect the check for the change to the device copy, and only now the data_t is only released twice if there is a device copy.
| @@ -2342,7 +2333,17 @@ parsec_device_kernel_epilog( parsec_device_gpu_module_t *gpu_device, | |||
|
|
|||
| assert( 0 <= gpu_copy->readers ); | |||
|
|
|||
| if( gpu_task->pushout & (1 << i) ) { | |||
| if (cpu_copy->flags & PARSEC_DATA_FLAG_DISCARDED) { | |||
| PARSEC_DEBUG_VERBOSE(20, parsec_gpu_output_stream, | |||
There was a problem hiding this comment.
Here is where the race-condition could happen.
There was a problem hiding this comment.
I think this is not a big issue here since we're not modifying the flags and neither the data_t nor the host copy will be freed, since the device copy maintains its reference. The worst thing that can happen is that we miss the flag being set and the release is delayed until we try to evict.
| parsec_atomic_lock( &gpu_copy->original->lock ); | ||
| /* get the next item before altering the next pointer */ | ||
| item = (parsec_list_item_t*)item->list_next; /* conversion needed for volatile */ | ||
| if( 0 == gpu_copy->readers ) { | ||
| if (cpu_copy->flags & PARSEC_DATA_FLAG_DISCARDED) { | ||
| parsec_list_item_ring_chop((parsec_list_item_t*)gpu_copy); |
There was a problem hiding this comment.
Same as above. The flag isn't changed and the device copy maintains a reference on the data_t so if we miss the update of the flags here we will evict into the host copy and then release everything.
Also OR the flag instead of assigning it. Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>
devreal
force-pushed
the
parsec-data-discard
branch
from
ba7ea19
to
fc1c247
Compare
Add a function
parsec_data_discardthat releases the data such that the host copy remains intact but does not prevent destruction of the data once all device copies have been released. This keeps the host copy available for device copies to inspect and avoids potential race conditions in the release process. During an eviction, copies of data with a discarded host copy are not transfered but put directly into the lru.Replaces some duplicated code with a call to
parsec_device_release_gpu_copy.