EMC blows CIFS benchmark away
But shies away from NFS
EMC has blown a CIFS benchmark away with a result 2.7 times better than the previous record, but why is it bothering? The only other suppliers on the list are Apple, Fujitsu and Silicon Graphics.
The benchmark is the SPECsfs2008 CIFS file access benchmark. There is a similar benchmark for NFS which is much more popular in terms of supplier submissions, with BlueArc, HP, Isilon, NetApp and others reporting results.
These two SPECsfs2008 benchmarks replaced an earlier SPEC SFS97 benchmark. Vendors have been updating their benchmarks to the 2008 standard and this is EMC's first reported result. We might hope for NFS ones to follow, as EMC reported several SPEC SFS97 results.
The previous best CIFS result was an Apple Xserve in early 2009 with the Snow Leopard version of Mac OS X. It achieved a throughput score of 44,347 operations/second, using 65 hard disk drives with a 9.1TB capacity. The overall response time was 1.89msecs. EMC's Celerra Gateway NS-G8 product, a filer head in a 3-node cluster set-up front-ending a drive array, achieved 118,463 on the throughput score, emphatically better, with an overall response time of 1.92msecs.
It did so using a 100 drive Symmetrix V-Max array. Generally we'd expect a Celerra gateway to sit in front of CLARiiON arrays, not the high-end enterprise Symmetrix. Looking at the system more closely, the 100 drives divide into 96 flash drives and four Fibre Channel disk drives. Oh, that's how EMC did it - flash drives are much faster than spinning rust.
There were 96 400GB STEC solid state drives (SSDs), 38.4TB of flash, and four 15K 400GB Seagate Cheetah disk drives, making 40TB of storage altogether. However, the HDDs were reserved for Celerra system use with all data filesystems residing on the flash. It was "divided into 48 2-disk RAID1 pairs, each with 4 logical units per drive, exported as 192 logical volumes." There were eight filesystems and each filesystem was striped across 48 disks.
This looks a pretty expensive configuration. It will be fascinating to see if EMC produces a SPECsfs2008 NFS benchmark as that will enable a direct comparison with Avere's multi-tiered hot box filer accelerators.
Apple put its 9.1TB XServe/Snow Leopard kit into the SPECsfs2008 NFS benchmark and returned a throughput score of 18,784, 2.36 times less than the CIFS throughput score. As a crude comparison, let's suppose then that the EMC Celerra Gateway NS-8 set-up above would score proportionately the same in the NFS benchmark. That would give it 45,043 on the NFS benchmark, not enough to match the several 100,000-plus scoring vendors.
EMC looks to have a much tougher job on its hands if it wants to blow away the SPECsfs2008 NFS benchmark the same way it has done the CIFS one. ®
Re: Don't compare NFS vs CIFS results
Couldn't agree more. These sorts of "leaps" of mathematics always smack of people who have never been involved in any real testing. It's one thing to make a leap within a single product or a vendor's line of products. It's another to assume that because one vendor saw a particular result, others will be the same.
I'm a big fan, El Reg. But this is the worst analysis ever.
Don't compare NFS vs CIFS results
In the article:
...and returned a throughput score of 18,784, 2.36 times less than the CIFS throughput score. As a crude comparison, let's suppose then that the EMC Celerra Gateway NS-8 set-up above would score proportionately the same in the NFS benchmark.
From the SPEC website referenced in the article:
SPECsfs2008 results may only be compared to other SPECsfs2008 results for the same protocol. SPECsfs2008_cifs and SPECsfs2008_nfs.v3 are not comparable because they are generated using completely different workloads.
So, no, it's not a crude comparison. You can't just do maths and assume that the platforms perform identically with different protocols. Because none do. NFS (2/3) are much simpler protocols than CIFS and it can show on the throughputs on devices that support both (natively). Where non-native CIFS emulation (e.g. Samba) is involved I'd expect the difference to be orders of magnitude different.
EMC have posted a SPEC SFS2008 NFS Number
As far as I can tell it is the same configuration they used for their CIFS number. And interestingly it seems does not suffer the same differential between CIFS and NFS performance that Apple does.
Apple's number is surprising - given that OS X is based on BSD I would have expected it to do better with NFS than CIFS, not the other way around.