MSA2062 - 25Gbps iSCSI connectivity

[Rob Leadbeater]

Hi,

We have an MSA2062 iSCSI model (R0Q82A) which was purchased a little over a year ago.

Looking at the current QuickSpecs I see that there's now 25Gbps SFPs available, which weren't an option when we purchased the array.

Is it possible to simply swap out our existing 10Gbps SFPs for 25Gbps SFPs?

Cheers,

Rob

[JonPaul]

@Rob Leadbeater 
Yes,  you can swap the MSA 10Gb SFPs for the MSA 25Gb SFPs.  The 25Gb SFPs will downshift to 10Gb if required (switch port or HBA only capable of 10Gb)
As a possible cheaper option there are multiple 25Gb DAC and AOC cables available also listed in the QuickSpecs.  With SFPs also make sure the optical cables are capable of 25Gb which is available with OM3 and OM4 cables (which you should already be using)

[xaviermn]

I'm looking for a new MSA 2062 using iSCSI, but when I look at the MSA 2062 end-to-end performance figures in the QuicksSpecs document, I see that there are a lot of differences in terms of performance when comparing FC, SAS, and iSCSI configurations. One of the reasons is the slower speed of the host protocol:

- 16 Gbps bus speed in FC
- 12 Gbps bus speed on SAS
- 10 Gpbs link speed in iSCSI Other important aspect is de protocol eficiency. And we know that iSCSI on TCP is the worst of the three protocols.

So the dilemma is about using a 25 Gbps link with iSCSI. If we connect the hosts directly to MSA using full 25 Gbps kits, we really gain performance. Does anyone have experience with this? Are there any performance comparisons using a 25 Gbps link with iSCSI? It's really recommended to upgrade these to 25 Gbps.

I would appreciate any feedback if anyone has experience with this setup.

Thank!

[JonPaul]

@xaviermn 
Thanks for attaching the table from the quickspecs it seems to be a little out of date.  Here is a link to the latest QS with updated numbers. 
https://www.hpe.com/psnow/doc/a00094629enw.pdf

I will point out what I see in this table and let you make an informed decision.
Sequential READs:
iSCSI system states a maximum throughput of 8700MB/sec.  Note that the other Protocols have higher ability (12000MB/sec).  This indicates that the system has potential for more performance so what is the bottleneck.  If the ports are all 10GbE then each port will be capped at a little over 1000MB/sec. There are only 8 ports so  the bottleneck is the host ports running at 10Gb.  So it would be reasonable to expect that Sequential READs would see a bump in performance with updating to 25Gb.

Sequential WRITEs:
Doing the same across protocol comparison it would be reasonable to expect a small bump in performance but the limit is definitely not the host ports, which should have the same potential as READs.

Random READs:
Again the same comparison across protocols shows a potential small bump in performance.  But also looking at the maximum IOPS for an individual 10G port which would be limited by throughput == 10Gb convert to Bytes => ~1000 MB divide by block size (8k) => 125,000 IOPS to fill a 10Gb port with data.  So the potential IOPS with 8 ports is ~1M IOPS.  The host ports are not the bottleneck .... unlikely to unlock additional speed with a port speed increase.

Random WRITEs:
See above

So if your workload is heavily sequential READs then you should see performance increase.  For other workloads you will likely see some performance increase but you should not expect 2.5x increase which may be inferred by increasing individual port speed from 10Gb -> 25Gb.
On the otherhand if your configuration is port limited (further limiting sequential potential) you could see significant sequential performance increases.

All of the above also assumes that the backend (disks) has the ability to meet the maximum performance of the system.  You can get a good idea of configuration performance with the HPE MSA Sizer tools.   https://ninjaonline.ext.hpe.com/
It is recommended to use the AUTO tools when initially configuring a system but using the manual tool can show you the potential of an existing setup in either a single or dual pool configuration.

[xaviermn]

@JonPaulThanks. Good analysis.

 I will look at all the options.