Add NUMA directory with Java server-side case study#42
Open
dianef wants to merge 1 commit into
Open
Conversation
adgubrud
requested changes
Jul 17, 2026
adgubrud
left a comment
adgubrud
left a comment
Contributor
There was a problem hiding this comment.
This article was easy to read and informative! Just a couple of small things to fix.
| numactl --cpunodebind=2 --membind=2 java -Xms24g -Xmx24g ... | ||
| ``` | ||
|
|
||
| This approach provides even stricter control over NUMA placement and may yield incremental improvements over the configuration-only approach demonstrated here. |
Contributor
There was a problem hiding this comment.
I'm assuming since you don't disclose a specific extra uplift when using numactl, you don't have this data? If you do, I think it would be good to provide a ballpark estimate of how much extra performance this can yield.
|
|
||
| ### Memory Subsystem Latency Improvements | ||
|
|
||
| Performance monitoring data collected using Intel's performance counter monitoring tool during both runs reveals significant improvements in memory subsystem latencies for the NUMA-optimized configuration: |
Contributor
There was a problem hiding this comment.
I would suggest giving a couple of pointers to potential performance monitors. I think PerfSpect and VTune would be good callouts.
|
|
||
| ### Microarchitectural Efficiency: Top-Down Analysis | ||
|
|
||
| Intel's Top-Down Microarchitecture Analysis Method (TMA) provides insight into how CPU cycles are spent across different execution pipeline categories. Comparing the two configurations reveals improved microarchitectural efficiency: |
Contributor
There was a problem hiding this comment.
Another article recently included this link to explain TMA, which I think would be appropriate to add here as well.
| - Contention for memory controllers on oversubscribed NUMA nodes | ||
| - Cache coherency traffic across NUMA boundaries | ||
|
|
||
| The NUMA-optimized 3-JVM configuration removes these barriers, allowing cores to remain active and productive. The 90% utilization demonstrates that the system is now able to keep cores busy doing useful work rather than stalling in wait states. |
Contributor
There was a problem hiding this comment.
Suggested change
| The NUMA-optimized 3-JVM configuration removes these barriers, allowing cores to remain active and productive. The 90% utilization demonstrates that the system is now able to keep cores busy doing useful work rather than stalling in wait states. | |
| The NUMA-optimized 3-JVM configuration removes these barriers, allowing cores to remain active and productive. The 90% CPU utilization demonstrates that the system is now able to keep cores busy doing useful work rather than stalling in wait states. |
| - **3-JVM configuration:** 90% average CPU utilization | ||
| - **Improvement:** +14 percentage points | ||
|
|
||
| This increase reveals that the 2-JVM configuration was leaving significant compute capacity untapped. The 76% utilization indicates cores were frequently idle—not because the workload was light, but because threads were **blocked waiting** for memory operations, likely due to: |
Contributor
There was a problem hiding this comment.
Suggested change
| This increase reveals that the 2-JVM configuration was leaving significant compute capacity untapped. The 76% utilization indicates cores were frequently idle—not because the workload was light, but because threads were **blocked waiting** for memory operations, likely due to: | |
| This increase reveals that the 2-JVM configuration was leaving significant compute capacity untapped. The 76% CPU utilization indicates cores were frequently idle—not because the workload was light, but because threads were **blocked waiting** for memory operations, likely due to: |
|
|
||
| - **Peak Bandwidth:** 572.8 GB/s | ||
| - **Minimum Latency:** 116.95 ns (local NUMA access) | ||
| - **Remote Latency:** 135-163 ns (40% penalty for cross-NUMA access) |
Contributor
There was a problem hiding this comment.
Suggested change
| - **Remote Latency:** 135-163 ns (40% penalty for cross-NUMA access) | |
| - **Remote Latency:** 135-163 ns (15-40% penalty for cross-NUMA access) |
| - [scikit-learn](software/scikit-learn/README.md) | ||
| - [MySQL & PostgreSQL](software/mysql-postgresql/README.md) | ||
| - [NUMA](software/NUMA/README.md) | ||
| - [Case Studies](software/NUMA/README.md) |
Contributor
There was a problem hiding this comment.
Can you please create another README in the software/NUMA/case_studies directory and then update this link to that README? I think just a table of contents in this new README will be enough.
As it is, it just links to the NUMA readme, which could be cleaner from a user experience perspective.
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
2 participants
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 PR adds content requested at Board of Advisors in Fall 2025. Compares performance of a Java workload that has been configured correctly to match the underlying NUMA configuration against one that ignores the NUMA configuration. It also creates a NUMA section where additional NUMA content can be added by future contributors.