Avere Systems has published some benchmark results for their FXT series of NAS acceleration products. At first glance, these look pretty good; the figures are based on the SPECsfs2008 benchmark. They achieved just over 20.000 IOs per second per node, with near-linear scaling of these results up to a 6-node cluster. The latter, according to Avere, “achieved a record-setting combination of 131,591 ops/sec throughput and minimal latency of 1.38ms ORT (overall response time).”
Unfortunately for them, SPEC publishes very complete reports about the tests performed and the exact hardware that was used for the tests. The NFS results are available here; the list immediately shows that all other products are “traditional” NFS servers, not caching products. It also shows the Avere system is not the fastest listed; but the only faster system uses 960 disks, so that’s not quite a fair comparison.
The benchmarks for the Avere system itself, however, show some rather interesting details. The Avere FXT units are meant as caches between the NFS fileserver and the clients, which should automatically detect “data hotspots” and cache those. In a real-world situation, the actual workload will be a mix of cached and non-cached data.
The test setup was carefully chosen to make sure all data on the backend fileserver fit into the FXT cache. In the single-node and 2-node cluster configurations, disks were removed from the fileserver to keep its capacity below the total cache size provided by the FXT. In addition to that, the quote “combination of 131,591 ops/sec throughput and minimal latency of 1.38ms ORT” is a bit misleading. According to the actual benchmark results available here, the response time was 4.6ms when reaching the 131,591 IOPS; which is significantly higher.
Still, the performance figures are impressive; if you’re looking for a way to make your NAS system faster, you should definitely consider evaluating the Avere products.
Related posts:
Your RSS reader
Daily e-mail
Twitter
What does this mean?
"…the list immediately shows that all other products are “traditional” NFS servers, not caching products."
Several tested systems use more DRAM and deliver far less performance.
Re: "…1.38ms ORT” is a bit misleading. According to the actual benchmark results available here, the response time was 4.6ms when reaching the 131,591 IOPS; which is significantly higher."
That's how this benchmark works — this is true of ALL SPEC results. When SPEC says "ORT", it stands for "Overall Response Time" and it's consistent across all reported results across all vendors.
Consider…competitive systems need 600-900 spindles to get this many IOPS. Think dollars-per-IOPS and power-per-IOPS! Bottom line, 131K SPEC-ops using only 79 spindles is absolutely stunning.
And they do it without Flash!!!