Speed & jumbo frames

iSCSI related applications
Post Reply
Veewan
New here
Posts: 9
Joined: Fri Mar 19, 2010 5:38 am

Speed & jumbo frames

Post by Veewan » Sun Mar 21, 2010 1:36 am

Ok I've been doing some testing with this box and I'm a bit concerned about the speeds I'm seeing. Unit is a brand new TS-459Pro with three WD 1TB drives (the new ones with 64MB cache and 6GB/s SATA) in RAID5.

With a 10GB LUN connected to a Server 2008R2 box, I've been trying various things. The most basic test was a straight copy of a single 10GB file from the SAN to the local disk which should produce the maximum throughput. This runs at around 60MB/s. Whilst acceptable, I was expecting a bit more from the reviews I'd read of the device.

I've also done a few tests using IOMeter. Running with huge 4MB requests at 100% read produces around 55MB/s, very close to the above file copy test. Obviously reducing the request size and read percentage will cause this to drop but I'm not sure by how much. I've just run a test with 4K requests at 75% read & 25% write and I'm seeing 5.4Mb/s read and 1.9Mb/s writes. These seem very low to me or am I wrong?

I've also tried enabling jumbo frames on both the QNAP and the server's Broadcom NIC but it seems to make very little difference. I've confirmed that jumbo frames are working by pinging the QNAP with an 8972 packet size with the DF flag set and this is succesful but the IOMeter results are virtually unaffected. Bizarrely, the straight file copy actually drops in speed to around 40-45Mb/s with jumbo frames enabled. This can't be right, can it?

Any and all advice appreciated as I'm a bit new to this iSCSI malarky :)

Thanks,
Toby.

PiroNet
Easy as a breeze
Posts: 295
Joined: Tue Dec 08, 2009 10:51 pm

Re: Speed & jumbo frames

Post by PiroNet » Sun Mar 21, 2010 6:01 am

Veewan wrote:I've also tried enabling jumbo frames on both the QNAP and the server's Broadcom NIC but it seems to make very little difference. I've confirmed that jumbo frames are working by pinging the QNAP with an 8972 packet size with the DF flag set and this is succesful but the IOMeter results are virtually unaffected. Bizarrely, the straight file copy actually drops in speed to around 40-45Mb/s with jumbo frames enabled. This can't be right, can it?


It is likely that you won't see any read/write improvement with jumbo frame on your setup, it will just decrease CPU usage!
60MB/s is a 'good' score but most important is the IOPS you get...
Some readings that might help you:
-http://deinoscloud.wordpress.com/2009/12/30/benchmark-tools-what-i-use/
-http://deinoscloud.wordpress.com/2009/09/11/understanding-disk-iops/

Cheers,

Veewan
New here
Posts: 9
Joined: Fri Mar 19, 2010 5:38 am

Re: Speed & jumbo frames

Post by Veewan » Sun Mar 21, 2010 11:47 pm

Thanks for that. I've just run IOMeter using the five recommended tests on your site and got the following:

8K, 50/50 r/w, 50/50 random: 132 IOps
64K 100 r, 100 sequential: 639 IOps
64K 100 w, 100 sequential: 494 IOps
256K 100 r, 100 sequential: 204 IOps
256K 100 w, 100 sequential: 163 IOps

Are these figures any good? If you can send me the sample IOMeter config file you mention on the site I'm happy to give that a try :)

Cheers,
Toby.

PiroNet
Easy as a breeze
Posts: 295
Joined: Tue Dec 08, 2009 10:51 pm

Re: Speed & jumbo frames

Post by PiroNet » Mon Mar 22, 2010 12:15 am

Veewan wrote:Thanks for that. I've just run IOMeter using the five recommended tests on your site and got the following:

8K, 50/50 r/w, 50/50 random: 132 IOps
64K 100 r, 100 sequential: 639 IOps
64K 100 w, 100 sequential: 494 IOps
256K 100 r, 100 sequential: 204 IOps
256K 100 w, 100 sequential: 163 IOps

Are these figures any good? If you can send me the sample IOMeter config file you mention on the site I'm happy to give that a try :)

Cheers,
Toby.


Check out the attached file. Rename to.ICF and in IOMeter open this new Test Configuration file.

The figures above are in the average, 494 IOPS at 64K is not too bad considering the trade-off on performance with RAID5.
If you're looking for performance go for RAID1 (eventually RAID10 if available). The trade-off here is space, you lose 50% for data mirroring and stripping.

BTW avoid RAID5 with more than 3 'desktop' drives, probability a second drives fails whilst rebuilding the failed disk is very important starting already at 4 disks!

Rgds,
You do not have the required permissions to view the files attached to this post.

Veewan
New here
Posts: 9
Joined: Fri Mar 19, 2010 5:38 am

Re: Speed & jumbo frames

Post by Veewan » Mon Mar 22, 2010 12:41 am

Ok thanks. Provided performance is "normal" then I'm happy - the device isn't going to be put under huge load.

As for drives, yeah not planning on adding a fourth drive but, if performance starts being an issue then I'll get another 1TB and reconfigure for RAID1 :)

Veewan
New here
Posts: 9
Joined: Fri Mar 19, 2010 5:38 am

Re: Speed & jumbo frames

Post by Veewan » Mon Mar 22, 2010 1:39 am

Incidentally, how much speed improvement (both bandwidth & IOps) could I expect by going with RAID1 or RAID10 over RAID5? I opted for the 459 over the cheaper 439 due to the increased processing power as I expected this to help keep RAID5 speed fairly close to RAID1 or 10 - is this not the case?

If there's a major speed improvement to be had then maybe I'll look at getting a fourth disk, but I don't want to do this for a minimal improvement.

Thanks

PiroNet
Easy as a breeze
Posts: 295
Joined: Tue Dec 08, 2009 10:51 pm

Re: Speed & jumbo frames

Post by PiroNet » Mon Mar 22, 2010 6:11 am

Veewan wrote:Incidentally, how much speed improvement (both bandwidth & IOps) could I expect by going with RAID1 or RAID10 over RAID5? I opted for the 459 over the cheaper 439 due to the increased processing power as I expected this to help keep RAID5 speed fairly close to RAID1 or 10 - is this not the case?

If there's a major speed improvement to be had then maybe I'll look at getting a fourth disk, but I don't want to do this for a minimal improvement.

Thanks


There is no secret, theoretically RAID5 and 6 are much slower than RAID1 or 10 due to the parity overhead that could be mitigated with big cache on RAID controllers and dedicated processors/ASICS but you find that on enterprise level storage most of the time. No need to say they are much more expensive than a QNAP :)

Many SOHO NAS devices on the market are missing these big R/W caches and enterprise class RAID controllers to mitigate performance hits of RAID5/6 configurations with 'desktop' HDDs, don't get me wrong your 459 is a fantastic product but not a Formula One car...yet!

Post Reply

Return to “iSCSI – Target & Virtual Disk”