[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <cc58c09e-bbb5-354a-2030-bf8ebb2adc86@iogearbox.net>
Date: Tue, 1 Jun 2021 08:57:37 +0200
From: Daniel Borkmann <daniel@...earbox.net>
To: robin.murphy@....com, jroedel@...e.de
Cc: Jussi Maki <joamaki@...il.com>, netdev@...r.kernel.org,
bpf <bpf@...r.kernel.org>, intel-wired-lan@...ts.osuosl.org,
davem@...emloft.net, anthony.l.nguyen@...el.com,
jesse.brandeburg@...el.com, hch@....de,
iommu@...ts.linux-foundation.org, suravee.suthikulpanit@....com
Subject: Re: Regression 5.12.0-rc4 net: ice: significant throughput drop
[ ping Robin / Joerg, +Cc Christoph ]
On 5/28/21 10:34 AM, Jussi Maki wrote:
> Hi all,
>
> While measuring the impact of a kernel patch on our lab machines I stumbled upon
> a performance regression affecting the 100Gbit ICE nic and bisected it
> from range v5.11.1..v5.13-rc3 to the commit:
> a250c23f15c2 iommu: remove DOMAIN_ATTR_DMA_USE_FLUSH_QUEUE
>
> Both recent bpf-next (d6a6a55518) and linux-stable (c4681547bc) are
> affected by the issue.
>
> The regression shows as a significant drop in throughput as measured
> with "super_netperf" [0],
> with measured bandwidth of ~95Gbps before and ~35Gbps after:
>
> commit 3189713a1b84 (a250c23^):
> $ ./super_netperf 32 -H 172.18.0.2 -l 10
> 97379.8
>
> commit a250c23f15c2:
> $ ./super_netperf 32 -H 172.18.0.2 -l 10
> 34097.5
>
> The pair of test machines have this hardware:
> CPU: AMD Ryzen 9 3950X 16-Core Processor
> MB: X570 AORUS MASTER
> 0a:00.0 Ethernet controller [0200]: Intel Corporation Ethernet
> Controller E810-C for QSFP [8086:1592] (rev 02)
> Kernel config: https://gist.github.com/joamaki/9ee11294c78a8dd2776041f67e5620e7
>
> [0] https://github.com/borkmann/stuff/blob/master/super_netperf
>
Powered by blists - more mailing lists