Repository navigation
hipblaslt-bench performance improvements - #1202
Merged
Merged
Conversation
…ocManaged or hipHostMalloc
pghysels
requested review from
Jinp800125,
KKyang,
Serge45,
TonyYHsieh,
aazz44ss,
hcman2,
imcarsonliao,
jichangjichang,
solaslin and
vin-huang
as code owners
October 3, 2024 20:16
Collaborator
|
Do we have the execution time of hipblaslt-test with/wo this PR? |
Contributor
Author
|
With these changes while the develop branch took On an MI210, 16 OpenMP threads. |
bethune-bryant
previously approved these changes
Oct 5, 2024
Contributor
In my testing it basically reduces total benchmark time to the sum of all kernel execution times, effectively removing all overhead. |
Contributor
Author
|
Sorry, I had to make a few small fixes. But I think this is ready now. |
bethune-bryant
previously approved these changes
Oct 7, 2024
Contributor
|
@jichangjichang @KKyang |
Contributor
Author
|
Sorry, I got confused about |
Contributor
Author
I can try to address execution time of hipblaslt-test in a follow up PR. |
bethune-bryant
approved these changes
Oct 11, 2024
cmingch
pushed a commit
to cmingch/hipBLASLt
that referenced
this pull request
Nov 13, 2024
* Add a memory pool, to reuse hip memory, either from hipMalloc/hipMallocManaged or hipHostMalloc * Initialize A, B, C on the device
tony-davis
added a commit
to ROCm/rocm-libraries
that referenced
this pull request
Nov 6, 2026
… them (#14608) JIRA ID : AIHPBLAS-5352 ## Motivation `hipblaslt-test`, the hipBLASLt test client, keeps the host buffers that each test frees in a pool, so the next test can reuse them instead of allocating again. These buffers are pinned (allocated with `hipHostMalloc`, so the operating system can neither page them out nor reclaim them), and the pool kept every one until the process exited. On the gfx1250 hardware test host, which has about 499 GiB of memory, the 93 cases with 2^31- and 2^32-element sizes enabled in #13504 left about 123 GiB pinned in one process. With 4 CI runners each splitting the suite into 4 parallel test processes, the out-of-memory killer took the runners down on 10-23 and 10-27. Separately, a failed host allocation returns a null buffer, and the test then writes through it. I don't think this stack guarantees the gfx1250 host never runs short. What it does is make the client hold much less pinned memory, give it back sooner, and skip a case with the amounts instead of getting killed, which is the part of the problem the client can fix. If pressure comes back, the 110 GiB container caps from #14569 keep a kill inside one job, and putting the lock back (#14982 removes it) is a one-file revert. So I'm not planning a long soak test before merging. The pool comes from ROCm/hipBLASLt#1202, which reused allocations to speed up `hipblaslt-bench` across about 500,000 small matmuls. Reuse still matters for large buffers: on the test host below, `hipHostMalloc` takes about 0.18 s per GiB and `hipHostFree` about 0.06 s per GiB. ## Technical Details - A pooled host buffer of 1 GiB or more (`kHostPoolRetainLimit` in `d_vector.hpp`) now stays cached only while consecutive tests reuse it. The pool records the test count when a buffer comes back, and `HostPoolReleaseListener`, a new hook that Google Test calls after every test, frees the large buffers that the test just ended did not use. A run of same-shape cases, such as the four transpose combinations (NN, NT, TN, TT) of one size, still shares one allocation. - When a test asks for a large buffer that no cached buffer fits, the pool first frees the large buffers that earlier tests left, then pins new memory. This keeps the peak near one test's footprint when the shape changes. - A host buffer no longer gets the pool's 20% extra capacity when the extra would take it to 1 GiB or more (`host_pool_capacity()`), so a buffer under the limit stays under it. The existing cutoff for skipping the extra is a quarter of device memory, so on gfx1250 nearly every host buffer got it. - `HipHostBuffer::memcheck()` reports a failed pinned allocation, and `testing_matmul` checks every host buffer after allocating, including the four temporary buffers a swizzled A or B is built in. A failure now skips the test with the requested size ("could not allocate N GiB of pinned host memory"). - A pooled host buffer under 1 GiB is released once 100 consecutive tests have not reused it (`kHostPoolSmallIdleTests`). The pool reuses a buffer only for requests between half its size and its size, so without this a run of shrinking requests kept every earlier buffer pinned. - In `hipblaslt-bench`, a failed host allocation now exits with a failure, as a failed device allocation already did, instead of reporting success. - Device memory behaves as before. `hipblaslt-bench` does not install the listener, so it never frees large buffers early; it does see the 20% change for host buffers of 1 GiB or more. The hipBLASLt library itself is unchanged. This is the first of four stacked PRs for AIHPBLAS-5352. #14609 makes the client's host-memory query respect container memory limits, #14610 runs the memory preflight for every matmul case, and #14982 drops the host-wide lock in the gfx1250 workflow. ## Test Plan - `HostMemory_pre_checkin` (new, `clients/tests/src/host_memory_gtest.cpp`, 6 tests; the `_pre_checkin` suffix puts it in the standard tier the PR checks run). The tests check that a large buffer survives one test boundary and is freed after the next, that a same-shape test reuses it (what the first test wrote is still there), that a large miss frees two buffers earlier tests left, that a 1 GiB request gets exactly 1 GiB, that a small buffer is released after 100 tests that do not reuse it (and kept when one does), and that a failed allocation leaves a null buffer whose size is reported. A 0.9 GiB request gets exactly 0.9 GiB (the old rule gave it 1.08 GiB, over the limit). The tests skip, naming the buffer, on a host that cannot pin that much, and the reuse test skips on a Windows host where the pool frees its idle buffers before every request. A temporary variant without the small-buffer aging failed the new test. - To confirm the tests can fail, I built two temporary variants of the pool. One undid four behaviors at once (no release at the end of a test, no release on a miss, the 20% extra restored, `memcheck()` always succeeding), and it failed 4 of the 5 tests. The other freed large buffers whether or not a test used them, and it failed 3. Every test failed under at least one variant. - The 82 large-memory cases that apply to gfx90a (the five name patterns from #14545), and the `smoke` and `quick` tiers, on develop and on this branch, for wall time and peak resident memory. - Handle leaks: #14714 and #14700 on develop give the matrix layouts, matmul descriptors and HIP events owners, so an early return no longer leaks them. To check that with this stack's new early returns, a preloaded library counted creates and destroys of all three, on the full stack, over 1,892 matmul cases, once normally and once under a 1 MiB host limit so that most cases skip at the preflight. - Adjacent: the `FastCheck_pre_checkin` and `FastCheckDevice_pre_checkin` suites, including `memory_shortfall_reclaims_idle_buffers`, which uses the pool's existing `ReleaseCached()`. `hipblaslt-bench` was built, because it compiles the non-gtest form of the new check, but not run. ## Test Result One MI210 (gfx90a) on a shared host with 2.2 TB of memory. This branch's clients were built against the 10-25 TheRock nightly hipBLASLt gfx90a device library. Each pair of runs went back to back, but the host was busy (load average near 300), so the times carry some noise. | Measurement | develop | this PR | | --- | --- | --- | | 82 large-memory cases, wall time | 316.9 s | 313.8 s | | 82 large-memory cases, peak resident memory | 81.9 GiB | 49.2 GiB | | 82 large-memory cases, mean resident memory | 66.9 GiB | 16.6 GiB | | Resident memory after the last large case | 74 GiB | 2 GiB | | `smoke` + `quick` (9,171 tests), two runs | 66.6 s, 62.0 s | 63.6 s, 64.1 s | - Large-memory cases: 82 of 82 passed on both. `smoke` + `quick`: 9,169 passed and 2 skipped (`GroupedMatmul.smoke_Fp16NN` and `smoke_Bf16NN`, which skip on this GPU) on both. - `HostMemory_pre_checkin`, `FastCheck_pre_checkin` and `FastCheckDevice_pre_checkin`: 64 of 64 passed. - A first version that freed every large buffer at the end of each test took 449.6 s for the large cases against 318.2 s on develop, which is why the pool keeps a buffer while the next test reuses it. - Handle leaks: 0 layouts, descriptors or events left, both normally (13,580 created) and with 409 cases skipped at the preflight (11,235 created). - With the small-buffer aging, `smoke` + `quick` passed 9,225 tests in 59.3 s with peak resident memory 2.18 GiB. - gfx1250: this PR alone passed the hardware workflow's hipBLASLt tests in [40934683157](https://github.com/ROCm/rocm-libraries/actions/runs/40934683157), and Archana dispatched the full stack in [41582794047](https://github.com/ROCm/rocm-libraries/actions/runs/41582794047), which passed hipBLASLt twice with no kills and no memory skips. Her run on the rebased stack, [41717825380](https://github.com/ROCm/rocm-libraries/actions/runs/41717825380), was clean too. All three kept the host-wide lock; #14982 covers the run without it. - Automated GitHub checks are pending. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests. ## Risk level Medium (3/5). Every `hipblaslt-test` case passes through the new listener and the host pool, so a mistake here would affect the whole suite. The hipBLASLt library and device behavior are unchanged; only the test and bench clients change. The bench sees the 20% change near 1 GiB, and now exits with a failure when a host allocation fails. ## Related - Work tracking: [AIHPBLAS-5352](https://amd-hub.atlassian.net/browse/AIHPBLAS-5352), fixes 1 (pool retention) and 4 (null host buffer) of its list. - Cause: #13504. CI workaround this work is meant to replace: #14545, merged on 10-28. - Where the pool came from: ROCm/hipBLASLt#1202. Windows already clears the pool on smaller-memory hosts for the same reason: #4793, #6121. ## Device / Architecture Coverage Host-side test-client code only, with no kernel, library or kernel-selection change, so it does not depend on the GPU architecture and the standard PR checks are enough. Measured on gfx90a. ## Flags / Guardrails None. The 1 GiB threshold is the constant `kHostPoolRetainLimit`. [AIHPBLAS-5352]: https://amd-hub.atlassian.net/browse/AIHPBLAS-5352?atlOrigin=eyJpIjoiNWRkNTljNzYxNjVmNDY3MDlhMDU5Y2ZhYzA5YTRkZjUiLCJwIjoiZ2l0aHViLWNvbS1KU1cifQ --------- Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This improves performance of hipblaslt-bench when timing a large number of matmuls.
There are 2 main changes
Both changes together significantly improve performance.
I have a test yaml input file with ~500K tests.
With these changes, the code runs in ~42 minutes, compared to 8 hours before (on MI300).
I initially used rocRAND, but then removed this dependency. I can add that back if it is preferred over my own naive random number generator.
This is based on a commit by @bethune-bryant