lkml.org 
[lkml]   [2013]   [Nov]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: Got iperf regression while intel_iommu is on, how to cut the cost of cache flushing
From
Eric,
I do think it is the IO_MMMU reason that brought such performance
down, while TSO/TSQ are not the way to cure such wounded performance.
I might misunderstand the purpose of TSO/TSQ---get rid of the
extra cache flushing pain of mapping and unmapping operation done by
Intel IOMMU. (sorry for last HTML mail, resend txt )

Ethan

On Tue, Nov 12, 2013 at 9:34 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> On Tue, 2013-11-12 at 09:03 +0800, Ethan Zhao wrote:
>> Eric,
>> We have tested the performance with the TSO and TSQ patches
>> merged, the result not good, even worse than kernel without those two
>> patches. any idea ?
>>
>> kernel : 3.11.x with TSO & TSQ merged. ( CONFIG_INTEL_IOMMU_DEFAULT_ON=y )
>> Network Interface : eth4
>> Network driver : be2net
>>
>> Average Bandwidth for :
>> 1.tcp-unidirectional test : 4385 Mbits/sec
>> 2.tcp-unidirectional-parallel: 9383 Mbits/sec
>> 3.tcp-bidirectonal test : 2755 Mbits/sec
>>
>> vs
>>
>> kernel : 3.11.x without TSO & TSQ patches.
>> (CONFIG_INTEL_IOMMU_DEFAULT_ON is not set)
>> Network Interface : eth4
>> Network driver : be2net
>>
>> Average Bandwidth for :
>> 1.tcp-unidirectional test : 7992 Mbits/sec
>> 2.tcp-unidirectional-parallel: 9403 Mbits/sec
>> 3.tcp-bidirectonal test : 5802 Mbits/sec
>>
>
> So it seems its not the TSO/TSQ changes, but
> CONFIG_INTEL_IOMMU_DEFAULT_ON being on instead of off.
>
>
>


\
 
 \ /
  Last update: 2013-11-12 03:41    [W:0.227 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site