Public bug reported:
== Comment: #0 - Carol L. Soto <cls...@us.ibm.com> - 2017-03-03 22:23:11 ==
---Problem Description---
Mellanox CX4/CX5 IB if running IPoIB in connected mode, it does not have the hw
checksum so when running IPoIB in power system the performance is low.
with iperf and 4 threads we get
[ 8] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48536
[ 4] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48538
[ 5] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48540
[ 6] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48542
[ 8] 0.0-10.0 sec 9.07 GBytes 7.79 Gbits/sec
[ 4] 0.0-10.0 sec 9.16 GBytes 7.87 Gbits/sec
[ 5] 0.0-10.0 sec 9.13 GBytes 7.84 Gbits/sec
[ 6] 0.0-10.0 sec 9.18 GBytes 7.88 Gbits/sec
[SUM] 0.0-10.0 sec 36.5 GBytes 31.4 Gbits/sec
if we add the following patches:
>From 68201fbbb04a030864f8560b05d43d8019f7f8df Mon Sep 17 00:00:00 2001
From: Michael Ellerman <m...@ellerman.id.au>
Date: Thu, 11 Aug 2016 16:03:14 +1000
Subject: powerpc/Makefile: Drop CONFIG_WORD_SIZE for BITS
>From b492f7e4e07a28e706db26cf4943bb0911435426 Mon Sep 17 00:00:00 2001
From: Paul Mackerras <pau...@ozlabs.org>
Date: Thu, 3 Nov 2016 16:10:55 +1100
Subject: powerpc/64: Fix checksum folding in csum_tcpudp_nofold and
ip_fast_csum_nofold
>From d4fde568a34a93897dfb9ae64cfe9dda9d5c908c Mon Sep 17 00:00:00 2001
From: Paul Mackerras <pau...@ozlabs.org>
Date: Thu, 3 Nov 2016 16:15:42 +1100
Subject: powerpc/64: Use optimized checksum routines on little-endian
then I can get with iperf
[ ID] Interval Transfer Bandwidth
[ 4] 0.0-10.0 sec 23.0 GBytes 19.7 Gbits/sec
[ 6] 0.0-10.0 sec 22.9 GBytes 19.7 Gbits/sec
[ 5] 0.0-10.0 sec 22.9 GBytes 19.7 Gbits/sec
[ 7] 0.0-10.0 sec 22.9 GBytes 19.6 Gbits/sec
[SUM] 0.0-10.0 sec 91.7 GBytes 78.7 Gbits/sec
Contact Information = Carol Soto/cls...@us.ibm.com
---uname output---
4.8.0-34-generic #36-Ubuntu SMP Wed Dec 21 17:22:53 UTC 2016 ppc64le ppc64le
ppc64le GNU/Linux
---Additional Hardware Info---
Need a power system and Mellanox CX4 IB card.
Machine Type = P8
---Steps to Reproduce---
just run iperf with IPoIB in connected mode and will see aroung 30Gbits/sec.
If we add the patches then we can get more than 70Gbits/sec.
** Affects: linux (Ubuntu)
Importance: Undecided
Assignee: Taco Screen team (taco-screen-team)
Status: New
** Tags: architecture-ppc64le bugnameltc-152252 severity-high
targetmilestone-inin---
** Tags added: architecture-ppc64le bugnameltc-152252 severity-high
targetmilestone-inin---
** Changed in: ubuntu
Assignee: (unassigned) => Taco Screen team (taco-screen-team)
** Package changed: ubuntu => linux (Ubuntu)
--
You received this bug notification because you are a member of Kernel
Packages, which is subscribed to linux in Ubuntu.
https://bugs.launchpad.net/bugs/1670247
Title:
Ubuntu 16.10: Network checksum fixes needed for IPoIB for Mellanox
CX4/CX5 card
Status in linux package in Ubuntu:
New
Bug description:
== Comment: #0 - Carol L. Soto <cls...@us.ibm.com> - 2017-03-03 22:23:11 ==
---Problem Description---
Mellanox CX4/CX5 IB if running IPoIB in connected mode, it does not have the
hw checksum so when running IPoIB in power system the performance is low.
with iperf and 4 threads we get
[ 8] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48536
[ 4] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48538
[ 5] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48540
[ 6] local 44.44.44.44 port 5001 connected with 44.44.44.45 port 48542
[ 8] 0.0-10.0 sec 9.07 GBytes 7.79 Gbits/sec
[ 4] 0.0-10.0 sec 9.16 GBytes 7.87 Gbits/sec
[ 5] 0.0-10.0 sec 9.13 GBytes 7.84 Gbits/sec
[ 6] 0.0-10.0 sec 9.18 GBytes 7.88 Gbits/sec
[SUM] 0.0-10.0 sec 36.5 GBytes 31.4 Gbits/sec
if we add the following patches:
From 68201fbbb04a030864f8560b05d43d8019f7f8df Mon Sep 17 00:00:00 2001
From: Michael Ellerman <m...@ellerman.id.au>
Date: Thu, 11 Aug 2016 16:03:14 +1000
Subject: powerpc/Makefile: Drop CONFIG_WORD_SIZE for BITS
From b492f7e4e07a28e706db26cf4943bb0911435426 Mon Sep 17 00:00:00 2001
From: Paul Mackerras <pau...@ozlabs.org>
Date: Thu, 3 Nov 2016 16:10:55 +1100
Subject: powerpc/64: Fix checksum folding in csum_tcpudp_nofold and
ip_fast_csum_nofold
From d4fde568a34a93897dfb9ae64cfe9dda9d5c908c Mon Sep 17 00:00:00 2001
From: Paul Mackerras <pau...@ozlabs.org>
Date: Thu, 3 Nov 2016 16:15:42 +1100
Subject: powerpc/64: Use optimized checksum routines on little-endian
then I can get with iperf
[ ID] Interval Transfer Bandwidth
[ 4] 0.0-10.0 sec 23.0 GBytes 19.7 Gbits/sec
[ 6] 0.0-10.0 sec 22.9 GBytes 19.7 Gbits/sec
[ 5] 0.0-10.0 sec 22.9 GBytes 19.7 Gbits/sec
[ 7] 0.0-10.0 sec 22.9 GBytes 19.6 Gbits/sec
[SUM] 0.0-10.0 sec 91.7 GBytes 78.7 Gbits/sec
Contact Information = Carol Soto/cls...@us.ibm.com
---uname output---
4.8.0-34-generic #36-Ubuntu SMP Wed Dec 21 17:22:53 UTC 2016 ppc64le ppc64le
ppc64le GNU/Linux
---Additional Hardware Info---
Need a power system and Mellanox CX4 IB card.
Machine Type = P8
---Steps to Reproduce---
just run iperf with IPoIB in connected mode and will see aroung 30Gbits/sec.
If we add the patches then we can get more than 70Gbits/sec.
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1670247/+subscriptions
--
Mailing list: https://launchpad.net/~kernel-packages
Post to : kernel-packages@lists.launchpad.net
Unsubscribe : https://launchpad.net/~kernel-packages
More help : https://help.launchpad.net/ListHelp
