#fdio-vpp: fdio vpp
Meeting started by DaveBarach at 16:00:57 UTC
(full logs).
Meeting summary
- CSIT (maciek reporting) (DaveBarach, 16:01:44)
- performance regression commentary (DaveBarach,
16:02:35)
- fix in 18.01.01 (DaveBarach,
16:02:48)
- report to be issued beginning of next
week (DaveBarach,
16:03:02)
- git bisect of release branch. Perf delta
exactly at branch pull, not clear why 18.01 differs by about 5% from
master (DaveBarach,
16:04:49)
- dpdk 17.11 artifacts posted in master / built
out of 18.01 branch are different sizes (DaveBarach,
16:05:15)
- no obvious root-cause from that exercise
(DaveBarach,
16:05:54)
- test indicates perf delta between POC and LF
infra. When built on local haswell machine, 18.01 identical to
master (DaveBarach,
16:06:55)
- DaveBarach - suggests disassembling code
between perf/non-perf binaries (alagalah,
16:08:24)
- @dwallacelf did digging into ELF files etc and
thought it might be a micro-architecture issue (alagalah,
16:09:16)
- Discussion about building locally (and building
DPDK from source) or in LF infra. (alagalah,
16:10:27)
- @mackonstan2 believes main issue is in NAT
eating 7k clocks ... mentioned the std dev is good (sic) but still
have performance gap (@alagalah - editorial - trying to follow what
@mackonstan2 is actually saying... if its where we are building it
or if its just NAT) (alagalah,
16:12:13)
- (alagalah,
16:12:50)
- @DaveBarach seems like NAT was enabling an
input node that wasn't needed until feature was enabled.
(alagalah,
16:13:14)
- Honeycomb (Marek reporting) (DaveBarach, 16:14:30)
- 18.01 released (DaveBarach,
16:14:36)
- Host stack (DaveBarach, 16:15:26)
- continuing integration w/ Envoy (DaveBarach,
16:15:38)
- compatibility w/ libevent2 (DaveBarach,
16:16:31)
- ACTION: @alagalah to
update Wiki to reflect a) how to use VCL and LDPRELOAD b) by
corollary of demonstrating that LDPRELOAD works with Envoy / Istio,
which actually use glibc (socket, bind etc) and libevent (read,
write) which uses epoll_ etc, its probably more accurate to say
VCL/LDPRELOAD supports glibc calls and to a large extent, libevent
(can list libevent calls that Envoy uses that we know we support
etc) .... (alagalah,
16:20:51)
- @dwallacelf points out that there are a large
number of set|getsockopts() we need to actually support which were
implemented to a use-case (Envoy/Istio) rather than the full gamut
of options (alagalah,
16:22:04)
- aarch64 (Tina reporting) (DaveBarach, 16:22:46)
- hoststack updates: refactor of builtin apps, rw
lock implementation that replaces the custom implementation in the
session layer, refactored segment manager and usage of valloc as a
central segment base va manager (florinc,
16:24:51)
- DHL can't find well-known shipping
address (DaveBarach,
16:25:24)
- 18.01 release (DaveBarach, 16:27:15)
- 18.01.01 release to deal e.g. w/ csit
performance issues, 4 or 5 commits (DaveBarach,
16:28:07)
- pls edit wiki w/ 18.04 release
deliverables (DaveBarach,
16:30:24)
- We should be putting these "new features" into
JIRA and put a link to that into the Release Deliverables...
(alagalah,
16:31:33)
- Ultimately thats the 18.04 RelMgr's call but
generally seen as a positive. dwallacelf speaking on behalf of
Chrisy (alagalah,
16:32:32)
- container build POC (DaveBarach, 16:33:59)
- hardware all set (DaveBarach,
16:34:04)
- no performance alarms in place (DaveBarach,
16:34:46)
- configure jobs to actually use the container
minions (DaveBarach,
16:35:39)
- merged patch to avoid container-unfriendly
manipulation of /proc (DaveBarach,
16:36:40)
- current JJB master either sends all streams to
current infra OR to POC (DaveBarach,
16:37:19)
- reluctant to throw big switch (DaveBarach,
16:37:47)
- two restructurings of JJB pending, to send less
than every job to container infra (DaveBarach,
16:38:25)
- ACTION: DaveW to
email list w/ cut-over date (DaveBarach,
16:44:36)
- transition to groups.io (alagalah,
16:45:20)
- need to ping LF, et al to figure out
status (DaveBarach,
16:45:42)
- DaveBarach : SCTP has a number of Coverity
defects ... who would be contact? (alagalah,
16:46:14)
- ACTION: florinc to
contact Marco (original submitter) (alagalah,
16:46:26)
- FOSDEM: Ray reports: (alagalah,
16:46:43)
- Jerome demo of vpp went great (DaveBarach,
16:47:04)
- full rooms for most presos, great feedback,
Highlight was Jerome Tollet's impromptu demo of VPP (I think with
TAPv2 40G ?) (alagalah,
16:47:08)
- interest in general for VPP + TCP (alagalah,
16:47:30)
- Brian Brooks on __sync builtins (DaveBarach, 16:49:00)
Meeting ended at 16:58:02 UTC
(full logs).
Action items
- @alagalah to update Wiki to reflect a) how to use VCL and LDPRELOAD b) by corollary of demonstrating that LDPRELOAD works with Envoy / Istio, which actually use glibc (socket, bind etc) and libevent (read, write) which uses epoll_ etc, its probably more accurate to say VCL/LDPRELOAD supports glibc calls and to a large extent, libevent (can list libevent calls that Envoy uses that we know we support etc) ....
- DaveW to email list w/ cut-over date
- florinc to contact Marco (original submitter)
Action items, by person
- alagalah
- @alagalah to update Wiki to reflect a) how to use VCL and LDPRELOAD b) by corollary of demonstrating that LDPRELOAD works with Envoy / Istio, which actually use glibc (socket, bind etc) and libevent (read, write) which uses epoll_ etc, its probably more accurate to say VCL/LDPRELOAD supports glibc calls and to a large extent, libevent (can list libevent calls that Envoy uses that we know we support etc) ....
- florinc
- florinc to contact Marco (original submitter)
People present (lines said)
- DaveBarach (33)
- alagalah (25)
- florinc (6)
- bbrian (4)
- collabot` (4)
- jdl (1)
- BillyM (1)
Generated by MeetBot 0.1.4.