Skip to content

Comments

Switch aarch64 pause to sb on armv9a+#2390

Closed
Nicoshev wants to merge 1 commit intofacebook:mainfrom
Nicoshev:export-D70250662
Closed

Switch aarch64 pause to sb on armv9a+#2390
Nicoshev wants to merge 1 commit intofacebook:mainfrom
Nicoshev:export-D70250662

Conversation

@Nicoshev
Copy link
Contributor

Summary:
SB (Speculation Barrier) is a modern barrier, mandatory from armv8.5a.
It achieves the same result as issuing DSB+ISB, but without having the cpu drop its instruction pipeline.

We have noticed 20% to 30% increased throughput, on the 16, 32 and 64 thread-count case within the small locks benchmark.

In the below results, 'Sum' is throughput:

before:

------- folly::MicroSpinLock 16 threads
Sum: 130891978 Mean: 1817944 stddev: 147111
Lock time stats in us: mean 1 stddev 33 max 14937
------- folly::MicroSpinLock 32 threads
Sum: 54681548 Mean: 759465 stddev: 105588
Lock time stats in us: mean 5 stddev 78 max 35925
------- folly::MicroSpinLock 64 threads
Sum: 24013546 Mean: 333521 stddev: 90571
Lock time stats in us: mean 11 stddev 179 max 90498

after:

------- folly::MicroSpinLock 16 threads
Sum: 169135465 Mean: 2349103 stddev: 227369
Lock time stats in us: mean 1 stddev 25 max 8463
------- folly::MicroSpinLock 32 threads
Sum: 67853388 Mean: 942408 stddev: 108821
Lock time stats in us: mean 3 stddev 63 max 17020
------- folly::MicroSpinLock 64 threads
Sum: 28845120 Mean: 400626 stddev: 61624
Lock time stats in us: mean 9 stddev 149 max 30879

Reviewed By: Gownta

Differential Revision: D70250662

Summary:
SB (Speculation Barrier) is a modern barrier, mandatory from armv8.5a.
It achieves the same result as issuing DSB+ISB, but without having the cpu drop its instruction pipeline.

We have noticed 20% to 30% increased throughput, on the 16, 32 and 64 thread-count case within the small locks benchmark.

In the below results, 'Sum' is throughput:

before:

------- folly::MicroSpinLock 16 threads
Sum: 130891978 Mean: 1817944 stddev: 147111
Lock time stats in us: mean 1 stddev 33 max 14937
------- folly::MicroSpinLock 32 threads
Sum: 54681548 Mean: 759465 stddev: 105588
Lock time stats in us: mean 5 stddev 78 max 35925
------- folly::MicroSpinLock 64 threads
Sum: 24013546 Mean: 333521 stddev: 90571
Lock time stats in us: mean 11 stddev 179 max 90498

after:

------- folly::MicroSpinLock 16 threads
Sum: 169135465 Mean: 2349103 stddev: 227369
Lock time stats in us: mean 1 stddev 25 max 8463
------- folly::MicroSpinLock 32 threads
Sum: 67853388 Mean: 942408 stddev: 108821
Lock time stats in us: mean 3 stddev 63 max 17020
------- folly::MicroSpinLock 64 threads
Sum: 28845120 Mean: 400626 stddev: 61624
Lock time stats in us: mean 9 stddev 149 max 30879

Reviewed By: Gownta

Differential Revision: D70250662
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70250662

Nicoshev added a commit to Nicoshev/folly that referenced this pull request Feb 26, 2025
Summary:

SB (Speculation Barrier) is a modern barrier, mandatory from armv8.5a.
It achieves the same result as issuing DSB+ISB, but without having the cpu drop its instruction pipeline.

We have noticed 20% to 30% increased throughput, on the 16, 32 and 64 thread-count case within the small locks benchmark.

In the below results, 'Sum' is throughput:

before:

------- folly::MicroSpinLock 16 threads
Sum: 130891978 Mean: 1817944 stddev: 147111
Lock time stats in us: mean 1 stddev 33 max 14937
------- folly::MicroSpinLock 32 threads
Sum: 54681548 Mean: 759465 stddev: 105588
Lock time stats in us: mean 5 stddev 78 max 35925
------- folly::MicroSpinLock 64 threads
Sum: 24013546 Mean: 333521 stddev: 90571
Lock time stats in us: mean 11 stddev 179 max 90498

after:

------- folly::MicroSpinLock 16 threads
Sum: 169135465 Mean: 2349103 stddev: 227369
Lock time stats in us: mean 1 stddev 25 max 8463
------- folly::MicroSpinLock 32 threads
Sum: 67853388 Mean: 942408 stddev: 108821
Lock time stats in us: mean 3 stddev 63 max 17020
------- folly::MicroSpinLock 64 threads
Sum: 28845120 Mean: 400626 stddev: 61624
Lock time stats in us: mean 9 stddev 149 max 30879

Reviewed By: Gownta

Differential Revision: D70250662
Nicoshev added a commit to Nicoshev/hhvm that referenced this pull request Feb 26, 2025
Summary:

X-link: facebook/folly#2390

SB (Speculation Barrier) is a modern barrier, mandatory from armv8.5a.
It achieves the same result as issuing DSB+ISB, but without having the cpu drop its instruction pipeline.

We have noticed 20% to 30% increased throughput, on the 16, 32 and 64 thread-count case within the small locks benchmark.

In the below results, 'Sum' is throughput:

before:

------- folly::MicroSpinLock 16 threads
Sum: 130891978 Mean: 1817944 stddev: 147111
Lock time stats in us: mean 1 stddev 33 max 14937
------- folly::MicroSpinLock 32 threads
Sum: 54681548 Mean: 759465 stddev: 105588
Lock time stats in us: mean 5 stddev 78 max 35925
------- folly::MicroSpinLock 64 threads
Sum: 24013546 Mean: 333521 stddev: 90571
Lock time stats in us: mean 11 stddev 179 max 90498

after:

------- folly::MicroSpinLock 16 threads
Sum: 169135465 Mean: 2349103 stddev: 227369
Lock time stats in us: mean 1 stddev 25 max 8463
------- folly::MicroSpinLock 32 threads
Sum: 67853388 Mean: 942408 stddev: 108821
Lock time stats in us: mean 3 stddev 63 max 17020
------- folly::MicroSpinLock 64 threads
Sum: 28845120 Mean: 400626 stddev: 61624
Lock time stats in us: mean 9 stddev 149 max 30879

Reviewed By: Gownta

Differential Revision: D70250662
@facebook-github-bot
Copy link
Contributor

This pull request has been merged in b22f4c5.

facebook-github-bot pushed a commit to facebook/hhvm that referenced this pull request Feb 27, 2025
Summary:
Pull Request resolved: #9591

X-link: facebook/folly#2390

SB (Speculation Barrier) is a modern barrier, mandatory from armv8.5a.
It achieves the same result as issuing DSB+ISB, but without having the cpu drop its instruction pipeline.

We have noticed 20% to 30% increased throughput, on the 16, 32 and 64 thread-count case within the small locks benchmark.

In the below results, 'Sum' is throughput:

before:

------- folly::MicroSpinLock 16 threads
Sum: 130891978 Mean: 1817944 stddev: 147111
Lock time stats in us: mean 1 stddev 33 max 14937
------- folly::MicroSpinLock 32 threads
Sum: 54681548 Mean: 759465 stddev: 105588
Lock time stats in us: mean 5 stddev 78 max 35925
------- folly::MicroSpinLock 64 threads
Sum: 24013546 Mean: 333521 stddev: 90571
Lock time stats in us: mean 11 stddev 179 max 90498

after:

------- folly::MicroSpinLock 16 threads
Sum: 169135465 Mean: 2349103 stddev: 227369
Lock time stats in us: mean 1 stddev 25 max 8463
------- folly::MicroSpinLock 32 threads
Sum: 67853388 Mean: 942408 stddev: 108821
Lock time stats in us: mean 3 stddev 63 max 17020
------- folly::MicroSpinLock 64 threads
Sum: 28845120 Mean: 400626 stddev: 61624
Lock time stats in us: mean 9 stddev 149 max 30879

Reviewed By: Gownta

Differential Revision: D70250662

fbshipit-source-id: 4148100a7368c6dc76c36e2ab88d4d2dabe88904
@AGSaidi
Copy link

AGSaidi commented Feb 28, 2025

Checked using SB with the lockhammer benchmarks and saw a perf improvement of 11-18%

@embg
Copy link

embg commented Feb 28, 2025

@AGSaidi is lockhammer OSS? if so, could you share your code change in lockhammer?

@AGSaidi
Copy link

AGSaidi commented Feb 28, 2025

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants