On 06 Sep 2013, at 4:10 PM, Ed Cashin <ecas...@coraid.com> wrote:

> I don't have a lot of experience with the other non-Coraid AoE targets that 
> are out there, but you might check whether one of them that's oriented more 
> toward performance could be useful to you.
> 
> That said, while checking the vblade README for the design goals, I noticed 
> that it advertises a capacity for 16 outstanding commands.  If you want to 
> try some tuning, you could adjust Bufcount in dat.h and then make sure your 
> settings in /proc are sufficient to allow the kernel to buffer 16 writes.  
> (Read commands are small.)

I ran vblade with -b to increase the buffer count and it improved performance 
quite a bit, but it's now maxing out the CPU. I found that bufcount above 64 
showed little or no improvement. There is however a big difference between 
using normal IO (dd with conv=fdatasync) and direct IO (dd with 
{o,i}flag=direct) on the initiator:

Test            MB/s      CPU    AvgPktSz  Direct MB/s   CPU     AvgPktSz
Disk Read       538       95%    2083         623        67%     4333
Disk Write      443       97%    2095         582        75%     4345
Ramdisk Read    655       97%    2083         778        69%     4333
Ramdisk Write   424      100%    2095         624        81%     4345

AvgPktSz shows the average packet size as measured by nettop. Wireshark 
confirms that "normal" IO generates 4132-byte packets while direct IO results 
in 8740-byte packets. I know Q 5.23 of the Coraid Linux FAQ says that AoE 
devices with an odd number of sectors result in 512-byte IO jobs, but mine have 
even sector counts. This is probably not the best way to benchmark but when I 
create a filesystem on top of my AoE device I get awful performance (50 MB/s) 
so there are obviously alignment issues.

Either way, looking at the CPU usage it's clear that vblade isn't going reach 
10 Gb/s.

I also tried other Linux targets:

kvblade: Doesn't compile against kernel 3.x.

ggaoed: About 25% slower than vblade:

Test            MB/s     CPU  Direct MB/s  CPU
Disk Read        446     71%     446       51%
Disk Write       355     63%     557       56%
Ramdisk Read     531     91%     627       67%
Ramdisk Write    399     85%     602       73%

qaoed: 25 - 50% slower than vblade:

Test            MB/s     CPU  Direct MB/s  CPU
Disk Read        282     77%     473       73%
Disk Write       259     85%     465       73%
Ramisk Read      291     99%     521       69%
Ramdisk Write    261     75%     467       75%

Unless I'm missing any further tuning options, none of the open source Linux 
AoE targets seem to be suitable for a 10 Gb/s SAN.

Regards,
Derick
------------------------------------------------------------------------------
How ServiceNow helps IT people transform IT departments:
1. Consolidate legacy IT systems to a single system of record for IT
2. Standardize and globalize service processes across IT
3. Implement zero-touch automation to replace manual, redundant tasks
http://pubads.g.doubleclick.net/gampad/clk?id=51271111&iu=/4140/ostg.clktrk
_______________________________________________
Aoetools-discuss mailing list
Aoetools-discuss@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/aoetools-discuss

Reply via email to