ORIGINALLY POSTED 3rd February 2010

12,777 views on developerworks

You may have already seen that SVC has yet again set a new world record benchmark when it comes to the Storage Performance Council’s SPC-1 benchmark.

The initial planning called for a full eight node test – as we normally do, and the manufacturing requests were put in to setup the 4000+ HDD configuration needed to saturate an 8 node CF8 cluster running the SVC 5.1. code. However, as this was mid 4Q, and customer orders are always fullfilled in preference to internal requests, late in the day we could only get hold of 2048 HDD in two DS8700 controller configurations.

Thus, we only enough backend storage to setup a four node benchmark. However, its worth noting that this four node cluster actually achieved a 15% increase in SPC-1 IOPs (315,043) compared with our previous eight node cluster. Thus proving that the new CF8 nodes pack more than double the throughput capability of the outgoing 8G4 nodes.

As we have shown in previous benchmarks, and the SVC performance white papers that are available through your sale team / business partner, SVC scales linearly as you add nodes. Thus, a full eight node cluster, with enough backend capability would be roughly double the performance of this four node result – the maths is left to the end user as en exercise so, but all in all an impressive result.

Since we had assembled enough SVC hardware to perform an eight node test, but were lacking in disk infrastructructure, the team running these tests in Tucson tried using a six node cluster. Mainly out of curiosity. Interestingly this resulted in an even higher world record number of 380,489 SPC-1 IOPs.

Two interesting conclusions can be drawn here. The first being that despite not having any more disk performance, SVC can squeeze just that little bit more out of the storage. Obviously the added cache capability of the nodes helped to push the number that bit higher. However, as our EMC friends try to tell us, SPC-1 will attempt to negate some caching effects, by moving the cache hit workloads around the target volumes… The second conclusion we can draw from this is that it negates what Chuck Hollis over at EMC has previously called out about SPC-1. That is, simply scaled with the number of HDD you had installed. Clearly this is not true.

I don’t want to open up the SPC-1 debate again – we all know where certain vendors (or should that be vendor) stand(s) – but their absence is clear – and if nothing else this has yet again proved useful to show the difference between our last generation, and next generation nodes.

There seems to have been confusion over the titles of these reports, and it may at first glance look more like this was a DS8700 result, whereas this was a combined SVC and DS8700 result. This has been raised with the team that published these results and the SPC and should be rectified soon.

These results clearly back up our statements regarding SVC’s scalability and its enterprise readyness – capable of sustaining more than four fully configured enterprise controllers. Users can start small at just two entry edition nodes, and with the new Virtual Disk System solution at a very competitve price. If your infrastrcuture is much larger, then the same enterprise ready functions can be used and scaled up to eight standard nodes per cluster.

---

The numbers quoted here are available on the Storage Performance Council website :
IBM SPC-1 Results
A00086 IBM System Storage DS8700 and IBM System Storage SAN Volume Controller v5.1 – 4 node
A00086 IBM System Storage DS8700 and IBM System Storage SAN Volume Controller v5.1 – 6 node

Comments from post on developerworks: