Re: Linux I/O subsystem performance (was: linuxcon 2010...)

From: Chris Worley
Date: Tue Aug 24 2010 - 16:31:45 EST


On Tue, Aug 24, 2010 at 11:43 AM, Vladislav Bolkhovitin <vst@xxxxxxxx> wrote:
> Pasi Kärkkäinen, on 08/24/2010 11:25 AM wrote:
>>
>> On Mon, Aug 23, 2010 at 02:03:26PM -0400, Chetan Loke wrote:
>>>
>>> I actually received 3+ off-post emails asking whether I was talking
>>> about initiator or target in the 100K IOPS case below and what did I
>>> mean by the ACKs.
>>> I was referring to the 'Initiator' side.
>>> ACKs == When scsi-ML down-calls the LLD via the queue-command, process
>>> the sgl's(if you like) and then trigger the scsi_done up-call path.
>>>
>>
>> Uhm, Intel and Microsoft demonstrated over 1 million IOPS
>> using software iSCSI and a single 10 Gbit Ethernet NIC (Intel 82599).
>>
>> How come there is such a huge difference? What are we lacking in Linux?
>
> I also have an impression that Linux I/O subsystem has some performance
> problems. For instance, in one recent SCST performance test only 8 Linux
> initiators with fio as a load generator were able to saturate a single SCST
> target with dual IB cards (SRP) on 4K AIO direct accesses over an SSD
> backend. This rawly means that any initiator took several times (8?) more
> processing time than the target.

While I can't tell you where the bottlenecks are, I can share some
performance numbers...

4 initiators can get >600K random 4KB IOPS off a single target...
which is ~150% of what the Emulex/Intel/Microsoft results show using 8
targets at 4KB (their 1M IOPS was at 512 byte blocks, which is not a
realistic test point) here:

http://itbrandpulse.com/Documents/Test2010001%20-%20The%20Sun%20Rises%20on%20CNAs%20Test%20Report.pdf

The blog referenced earlier used 10 targets... and I'm not sure how
many 10G ports per target.

In general, my target seems capable of 65% the local small-block
random write performance over IB, and 85% the local small-block
random read performance. For large block performance, ~95% efficiency
is easily achievable, read or write (i.e. 5.6GB/s over fabric, where
6GB/s is achievable on the drives locally at 1MB random blocks).
These small-block efficiencies are achievable only when tested with
multiple initiators.

The single initiator is only capable of <150K 4KB IOPS... but gets
full bandwidth w/ larger blocks.

If I were to chose my problem, target or initiator bottleneck, I'd
certainly rather have an initiator bottleneck rather than Microsoft's
target bottleneck.

> Hardware used for that target and
> initiators was the same. I can't see on this load why the initiators would
> need to do something more than the target. Well, I know we in SCST did an
> excellent work to maximize performance, but such a difference looks too much
> ;)
>
> Also it looks very suspicious why nobody even tried to match that
> Microsoft/Intel record, even Intel itself who closely works with Linux
> community in the storage area and could do it using the same hardware.

The numbers are suspicious for other reasons. "Random" is often used
loosely (and the blog referenced earlier doesn't even claim "random").
If there is any merging/coalescing going on, then the "IOPS" are
going to look vastly better. If I allow coalescing, I can easily get
4M 4KB IOPS, but can't honestly call those 4KB IOPS (even if the
benchmark thinks it's doing 4KB I/O). They need to show that their
advertised block size is maintained end-to-end.

Chris
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/