When trimming old allocations in usm reuse start from largest
allocations.
This will reduce memory usage more quickly once max hold time is hit.
Related-To: NEO-6893, NEO-14429
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
Real allocation size should be used to properly apply limits and allow
more usm reuse hits.
Related-To: NEO-6893
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
Save svmData on putting into reuse, instead of searching each time.
Change UNRECOVERABLE_IF to DEBUG_BREAK_IF.
Related-To: NEO-6893
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
Host usm and device usm for igfx checks system memory usage.
Device usm for dgfx checks local memory usage.
If used memory is above limit threshold:
- no new allocations will be saved for reuse
- cleaner will use shorter hold time of 2 seconds
- cleaner will free all eligible allocations, regardless of async
deleter thread having work
Motivation: in case of gfx memory being full, making resident new
allocations will require evictions which leads to massive slowdown on
enqueue calls.
This change aims to minimize cases where extra memory usage from usm
reuse mechanism leads to above situation.
Related-To: NEO-6893, NEO-14160
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
If flag "LogUsmReuse" is set, usm reuse will log operations to csv file.
Each line will contain: timestamp, host/device, operation type,
allocation size, true/false whether operation succeeded.
This data can then be used to produce graphs and help in analyzing
usm reuse in a particular workload.
Related-To: NEO-6893
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
Cleaner thread will run every 15ms instead of 2s.
Allocations will be held for at least 10s.
If deferred deleter has elements to release, will skip cleaning cache.
Will clean only 1 allocation per cache, per cleaning run.
Related-To: NEO-6893
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
mechanism for freeing allocations saved for reuse that have not been
used in a given time
Related-To: NEO-13425
Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
Addresses regressions from the reverted merge
of the tbx fault manager for host memory.
Recursive locking of mutex caused deadlock.
To fix, separate tbx fault data from base
cpu fault data, allowing separate mutexes
for each, eliminating recursive locks on
the same mutex.
By separating, we also help ensure that tbx-related
changes don't affect the original cpu fault manager code
paths.
As an added safe guard preventing critical regressions
and avoiding another auto-revert, the tbx fault manager
is hidden behind a new debug flag which is disabled by default.
Related-To: NEO-12268
Signed-off-by: Jack Myers <jack.myers@intel.com>
Addresses regressions from the reverted merge
of the tbx fault manager for host memory.
This fixes attempts by the tbx fault manager
to protect/unprotect host buffer memory, even
if the host ptr was not driver-allocated.
In the case of the smoke test that triggered
the critical regression, clCreateBuffer was
called with the CL_MEM_USE_HOST_PTR flag.
The subsequent `mprotect` calls on the
provided host ptr then failed.
Related-To: NEO-12268
Signed-off-by: Jack Myers <jack.myers@intel.com>
In TBX mode, the host could not write to host buffers after access from device
code due to the lack of a migration mechanism post-initial TBX upload.
Migration is unnecessary with real hardware, but required for TBX.
This patch introduces a new page fault manager type that extends the original
CPU fault manager, enabling automatic migration of host buffers in TBX mode.
Refactoring was necessary to avoid diamond inheritance, achieved by using a
template parameter as the base class for OS-specific fault managers.
Related-To: NEO-12268
Signed-off-by: Jack Myers <jack.myers@intel.com>
