Skip to content

test: benchmark memory optimizations - #324

Closed
l0wl3vel wants to merge 23 commits into
masterfrom
feat/memory-consumption-test
Closed

l0wl3vel wants to merge 23 commits into
masterfrom
feat/memory-consumption-test

Conversation

@l0wl3vel

Copy link
Copy Markdown
Contributor

Description

TODO

Used AI-Tools ✨

  • TOOL used for generation

l0wl3vel added 22 commits July 28, 2026 12:25
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
includes
sonic-net/sonic-platform-vpp#212
and sonic-net/sonic-platform-vpp#220
for troubleshooting reasons

Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Removed hacks, that just about worked in sonic-vs but broke in sonic-vpp
- sonic-vpp requires first time init to generate
  VPP config files from hwsku lanemap and port
  config. If skipped the syncd container, in which
  vpp runs, will crash immediately
- /etc/sonic/sonic-environment is generated on first boot from
  /usr/share/sonic/device/x86_64-kvm_x86_64-r0/default_sku,
  so writing it serves no purpose as we reenabled firstboot
- switch to telnetlib3, due to telnetlib being deprecated

Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
- Split pipelines to build sonic_vs and sonic_vpp images
- Rename sonic flavor to sonic_vs
- Add sonic-vpp as sonic_vpp flavor
- Run integration tests for sonic_vpp

Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
fixes sonic_vs with older (202505) image

Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
does nothing anymore

Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
it is configured using the sonic startup config, we run lldp.service, so
we do not need it anymore

Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
Signed-off-by: Benjamin Ritter <benjamin.ritter@x-cellent.com>
This reverts commit c9e1837.

Management VRF requires some more work around bootstrapping the switch
Every leaf and machine is a QEMU VM, and QEMU RSS behaves as a high-water
mark: the guest touches all of its RAM through the page cache eventually,
so a guest started with -m 4096 stays resident at ~4 GiB no matter how
little it actually needs. virtio-balloon with free page reporting barely
helps because the page cache keeps almost nothing on the free lists.

Add three opt-in knobs, all defaulting to the previous behaviour:

  MINI_LAB_LEAF_MEMORY / MINI_LAB_MACHINE_MEMORY  guest RAM, substituted
    into the topologies. All three launchers read QEMU_MEMORY (SONiC
    launch.py, machine launch.py and vrnetlab for the dell flavors).
  MINI_LAB_KSM      host kernel samepage merging, dedupes the identical
    guest RAM of leaf01/leaf02 and of the machine VMs
  MINI_LAB_THP      transparent hugepage policy

KSM and THP are host global and are applied/reverted by
scripts/memory-tuning.sh, which records the pristine values so the host
is left as it was found.

scripts/memory-profile.sh bundles the knobs into one-factor-at-a-time
profiles (baseline, low-memory, ksm, thp-madvise, all) so each knob can
be measured on its own. baseline pins KSM and THP explicitly instead of
leaving them untouched, so a comparison is not skewed by the host state.

scripts/memory-trace.py samples host and per-container memory plus QEMU
RSS into a CSV; scripts/memory-report.py renders per-run summaries and a
cross-run comparison. The integration test starts the tracer and stops it
in an EXIT trap, so a profile that is too tight for a flavor still
produces data instead of nothing.

The integration workflow builds its matrix from flavors x profiles. Pull
requests and pushes run baseline only so regular CI cost is unchanged;
workflow_dispatch defaults to the full matrix. Runs are serialised
because the host level numbers would otherwise be contaminated by
concurrent labs. Each run uploads its trace, and a final job merges them
into a per-flavor comparison table.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Archived in project

Development

Successfully merging this pull request may close these issues.

1 participant