Thanks Jesse and Dan - comments to both below. I will not be able to 
access this box until the weekend, so will gather the information 
requested and anything else that may be useful.

Regards,

Richard

Dan Williams wrote:
> [ adding Dave, Maciej ]
>
> On 10/4/2010 1:13 PM, Brandeburg, Jesse wrote:
>>
>> Dan might be able to further help with the ioatdma portion, this sounds
>> like a possible regression in 4.00 ioatdma?
>
> Smells that way.  I'm traveling internationally this week, but I can try
> to reproduce this on my 5400 next week when I'm back (unless Dave or
> Maciej have a chance to take a look).  Question below:
>
>>
>> On Sat, 2 Oct 2010, Richard Scobie wrote:
>>
>>> I have a Tyan S5396 5400 based board, which has been running
>>> 2.6.30.10-105.2.23.fc11.x86_64and has a PRO/1000 PT (82571EB) Quad
>>> adapter, along with two onboard 80003ES2LAN adapters.
>>>
>>> This has performed fine with the ioatdma loaded and the BIOS is the latest.
>>>
>>> dca service started, version 1.8
>>> ioatdma 0000:00:0f.0: PCI INT A ->   GSI 16 (level, low) ->   IRQ 16
>>> ioatdma 0000:00:0f.0: setting latency timer to 64
>>> ioatdma 0000:00:0f.0: Intel(R) I/OAT DMA Engine found, 4 channels,
>>> device version 0x20, driver version 3.64
>>>      alloc irq_desc for 62 on cpu 0 node 0
>>>      alloc kstat_irqs on cpu 0 node 0
>>> ioatdma 0000:00:0f.0: irq 62 for MSI/MSI-X
>>>      alloc irq_desc for 63 on cpu 0 node 0
>>>      alloc kstat_irqs on cpu 0 node 0
>>> ioatdma 0000:00:0f.0: irq 63 for MSI/MSI-X
>>>      alloc irq_desc for 64 on cpu 0 node 0
>>>      alloc kstat_irqs on cpu 0 node 0
>>> ioatdma 0000:00:0f.0: irq 64 for MSI/MSI-X
>>>      alloc irq_desc for 65 on cpu 0 node 0
>>>      alloc kstat_irqs on cpu 0 node 0
>>> ioatdma 0000:00:0f.0: irq 65 for MSI/MSI-X
>>> ioatdma 0000:00:0f.0: DCA is disabled in BIOS
>>
>> hm, I think this note above is critical that it is missing from the below.

Although I believe neither 82571EB or 80003ES2LAN make use of DCA. This 
board does not have an explicit "DCA Enable" feature - just Crystal 
Beach, which is enabled.

>>
>>> I have just updated to 2.6.35.7 stable and with ioatdma loaded, samba RX
>>> performance drops from normal levels, to about 500kB/s with the smbd
>>> process at 100% CPU.
>>
>> does netstat -s show it all to be from retransmits or SACKs?
>
> I assume smbd normally does not consume 100% cpu.  Can you send the
> output of:

No, without ioatdma, around 22%.

> perf record -a sleep 5; perf report
>
> ...while the bad performance is occurring.  That might give a clue as to
> what is tanking the performance.
>
>>
>>> dca service started, version 1.12.1
>>> ioatdma: Intel(R) QuickData Technology Driver 4.00
>>> ioatdma 0000:00:0f.0: PCI INT A ->   GSI 16 (level, low) ->   IRQ 16
>>> ioatdma 0000:00:0f.0: setting latency timer to 64
>>>      alloc irq_desc for 80 on node -1
>>>      alloc kstat_irqs on node -1
>>> ioatdma 0000:00:0f.0: irq 80 for MSI/MSI-X
>>>      alloc irq_desc for 81 on node -1
>>>      alloc kstat_irqs on node -1
>>> ioatdma 0000:00:0f.0: irq 81 for MSI/MSI-X
>>>      alloc irq_desc for 82 on node -1
>>>      alloc kstat_irqs on node -1
>>> ioatdma 0000:00:0f.0: irq 82 for MSI/MSI-X
>>>      alloc irq_desc for 83 on node -1
>>>      alloc kstat_irqs on node -1
>>> ioatdma 0000:00:0f.0: irq 83 for MSI/MSI-X
>>
>> some of the other messages above are missing too, like DMA engine found,
>> and driver version 3.64
>>
>>>
>>>
>>> If it is not loaded, performance returns to normal levels, but this
>>> network RX performance degradation is not seen with rsync.
>>
>> rsync and samba may be using different sockets sizes which may prevent
>> ioatdma from being utilized to do copy offload.  This might explain rsync
>> working.  Can you check the stats in
>>
>> cat /sys/class/dma/dma?chan?/memcpy_count
>>
>> and see if the stats don't increment for one or the other?
>>
>
> Similarly you can:
>
> echo $((256*1024))>  /proc/sys/net/ipv4/tcp_dma_copybreak
>
> ...to effectively disable dma.
>
> --
> Dan

------------------------------------------------------------------------------
Virtualization is moving to the mainstream and overtaking non-virtualized
environment for deploying applications. Does it make network security 
easier or more difficult to achieve? Read this whitepaper to separate the 
two and get a better understanding.
http://p.sf.net/sfu/hp-phase2-d2d
_______________________________________________
E1000-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/e1000-devel
To learn more about Intel® Ethernet, visit 
http://communities.intel.com/community/wired

Reply via email to