SPC(-2) performance report as of January 2021

In $/MBPS, ETERNUS, Fujitsu, Gluesys AnyStor-700EK, MBPS, OceanStor, SPC, SPC-2, TTA by AdministratorLeave a Comment

This Storage Intelligence (StorInt™) dispatch covers Storage Performance Council (SPC) results[1]. Since our last report on SPC-1, three months ago[2], there have been 1 new SPC-1 v3 submissions, the Huawei OceanStor™ 5110F Enhanced and one new SPC-2 submission, the TTA GlueSys AnyStor-700ED. Although the Huawei submission was a solid performer (740K IOPS and an LRT of 0.22 msec) at a decent price ($318.6K USD), alas it did not place on any of our top 10 charts. On the other hand the TTA submission did reach top 10 in one of our charts and deserves some discussion. 

SPC-2 results

We begin our discussion with top 10 SPC-2 Price-Performance™  or $/MBPS chart in Figure 1.

Figure 1 Top 10 SPC-2 $/MBPS 

With a ~11.0 GB/sec  and a price of $37.6K, the TTA GlueSys AnyStor-700ED was able to come in as our new #1 with a Price-Performance ($/MBPS) of $3.42 ($3.43 as reported by SPC). The GlueSys AnyStor-700ED came with dual controllers with a combined 64GB of cache,24 1.92TB SAS SSDs and used 4 16Gbs FC front end interfaces in what looks like a 4U rack mount solution. It didn’t seem like any special hardware was in the system. But it wasn’t that long ago where a 11.0GB/sec score would get you into the top 10 in MBPS.  

Next, we report on another chart we don’t see very often our SCI computed, (Geomean) average response time across the 3 workloads (LFP, LDQ and VoD) in Figure 2.

Figure 2 Top 10 SPC-2 LDQ-LFP-VoD composite average workload response times 

We don’t discuss SPC-2 response time (RT) much for throughput workloads as it doesn’t seem as important when a system is pushing large amounts of data. But, SPC-2 in their FDR does publish RT information for all three of their work loads and all one needs to do is 1) select which RT to use, 2) extract these from the FDR and 3) compute the (geomean) average. The new TTA GlueSys AnyStor-700ED came in at our new #10 with an composite average RT time of 0.66 msec. Not bad when one considers the average price of rest of the products in this top ten (~$440K), over ten times as much as the TTA GlueSys AnyStore-700ED system.

One might ask why throughput response times are important. Well in today’s AI/ML/DL workloads one doesn’t want to keep those expensive GPUs/TPUs/IPUs/MPUs waiting for data. And it seems at this price one could throw 10 of these systems to support an aggregate of 110GB/sec which would seem more than enough to train the new 1.6 trillion parameter Google’s language model[3]. But what do I know. 

We end our SPC-2 result review with a chart we’ve presented before. The TTA GlueSys didn’t place on this chart but it does show the state of the art (at least in SPC-2 submissions) in MBPS. So Figure 3 is our MB/sec performance for all three workloads (LFP, LDQ, & VoD) ranked by MBPS. 

Figure 3 SPC-2 Top 10 MB/sec showing each workload performance

As you may recall, SPC-2 MBPS is an (mean) average of the three workloads peak throughput rate. We often find it informative to see how the three workloads actually perform rather than just looking at the aggregate.  For example, the #1 Fujitsu ETERNUS DS 8890 S3 has a 70.1 GB/sec average across it’s three workloads bit did very well on the LDQ (84.1GB/sec) and not so well on LFP (52.6GB/sec).

Just as a comparison, the TTA GlueSys AnyStor-700ED did 11.0GB/sec in aggregate but did 9.8GB/sec, 12.1GB/sec and 11.0GB/sec in LFP, LDQ and VoD respectively. The Oracle ZFS ZS5-2 AFA only managed 20.0 in LFP and had an aggregate score of 24.4GB/Sec.  

Significance

New SPC-2 submissions have been infrequent at best of late. But the need for (small file) throughput is higher now than ever before with the new AI/ML/DL workloads coming out of the enterprise.  We would suggest that SPC add a new workload to their SPC-2 composite, one that emulates the small file throughput used for DL training and real time inferencing. That way we might see some more interest from storage vendors to submit and enterprise customers might be able to once again, use SPC-2 results as a criteria to judge storage systems they wish to purchase. 

As always, suggestions on how to improve any of our performance analyses are welcomed.

[This storage performance was originally sent out to our newsletter subscribers in January of 2021.  If you would like to receive this information via email please consider signing up for our free monthly newsletter (see subscription request, above right) and we will send our current issue along with download instructions for this and other reports. Dispatches are posted to our website at least a quarter or more after they are sent to our subscribers. ]

Silverton Consulting, Inc., is a U.S.-based Storage, Strategy & Systems consulting firm offering products and services to the data storage community


Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.