next: Changes queued for the v2016.07 release #956
Conversation
Set the shared memory path (shm.path) to a private namespace for each app with prefix "app/$name". This means that apps can create shm objects such as counters and by default these will appear in a local namespace for that app.
I rewrote the git-workflow explanation with the goal of briefly but clearly explaining both how to submit a change to Snabb Switch and also how to be a subsystem maintainer. This is intended to be a chapter in the manual.
Contains "XXX rewrite" where old sections have been removed but replacements not written yet.
This file now contains a diagram so it needs to have separate .src.md and .md versions. This can be cleaned up when these changes eventually merge with the on-demand markdown (#829).
also stopped capitalizing "Pull Request" because it seemed awkward.
- Use "apps/" instead of "app/" for uniformity - Set shm path to "apps/$name" when calling `app:stop' too - Unlink "apps/$name" after `app:stop' using `shm.unlink' - Add a test case to core.app selftest
# Conflicts: # src/core/app.lua
…eric representation.
Resolved conflicts with incoming changes from master branch: *.src.md replaced with simply *.md "Snabb Switch" being renamed to "Snabb"
Ditaa images are not kept in tree anymore...
option and support injecting a function to determine the current time.
This reverts commit 8bb3215.
MRG_RXBUF is enabled by default in the beginning, and QEMU will initially negotiate a feature set with Snabb NFV that includes MRG_RXBUF. This adds a field onto the virtio header, in legacy mode. However if we later negotiate to not have MRG_RXBUF, we need to re-set this value to not have the extra fields. (cherry picked from commit df8e83c)
|
I cherry-picked the virtio-net fix df8e83c from lwaftr repo for renegotiating virtio-net options with guests that switch drivers while running. |
…dicate." Reason: its actually slower than the initial naive version. This reverts commit c186591. # Conflicts: # src/apps/vhost/vhost_user.lua
|
@lukego I have tracked down the regression to the statistics counters in vhost_user. I pushed e3fcbbf to #931 which reverts a premature optimization I did, and this removes half of the slowdown. That leaves us with a 2.5% regression. Now looking into what can be done about the remaining slowdown. Edit: well “premature optimization” is the wrong term here really as I tried to outsmart the compiler and failed. |
|
Profiling tells me that 10% of time is spent in
|
|
Alternatively we could omit counting successfully processed rx/tx packets in vhost_user, and rely on the existing link statistics to reflect these. (rx/tx packes/bytes app-counters are mostly redundant with respect to link stats.) |
# Conflicts: # src/apps/vhost/vhost_user.lua # src/lib/protocol/README.md
Fixed packet documentation merge conflict
|
@eugeneia Looking at a new Hydra run it looks like the same performance regression is there on both @domenkozar I would like to generalize these CI jobs. Instead of specifically testing for a performance regression on the 2016.07 release they could be continuously showing the comparative performance of a collection of branches: Meta: I reckon that all "out of band" information we are discussing should be brought "in band" somehow. For example, if profiler information is required for interpreting results then the CI tests should include runs that are profiled and publish the reports. This way we are all looking at the same information and can see exactly how it was produced. |
|
@domenkozar This is really cool testing to have up and running, btw! I am especially impressed that Hydra is smart enough to reuse all previous results that are still valid. So if one branch is updated then its benchmarks will be rerun but the results for the unmodified branches will be reused. This seems extremely efficient and means we can include a lot of branches in the benchmark campaigns. |
|
@lukego I've renamed the CI jobs to Once By "profiler information" do you mean we should write those tests and include them? I'm +1 on that, let's add more regressions tests. |
The counter value for the interface speed should be in units of bps. The ifSpeed SNMP object must obey RFC3635 sec. 3.2.8.
|
Looks like SnabbBot errored out due to snabblab/snabblab-nixos#52 |
# Conflicts: # src/apps/vhost/vhost_user.lua
Allow multiple CPUs via --cpu argument
Changes queued up for v2016.07.
Fixes needed:
core.timeris printing tens of thousands of debug messages. Have to account for why this has started happening and likely remove some debug printouts (seems excessive to print a message on every timer invocation even when developer_debug is enabled... could be a use case for timeline instead soon).benchmark-reportjob.