SPEC sfs2014 (vda) performance report as of February 2021

In Quantum, SPECsfs, SPECsfs2014, StorNExt, VDA workload by AdministratorLeave a Comment

This Storage Intelligence (StorInt™) dispatch covers SPEC sfs2014 benchmarks[1]. There was one new vendor submission for the VDA (video data acquisition) workload over the last 6 months and that was the Quantum StorNext 7.0.1 with F-Series Storage. 

SPEC sfs2014_vda benchmark results

The SPEC sfs2014_vda workload simulates high-definition video data streams. The VDA workload consists of two activities, one is high-def, camera data streams (100% write) and the other is a set of companion applications (~84% random read). We could find no specification for the ratio of camera streams to application IO, but we believe it’s at least 50% stream IO with each stream representing ~36Gbps of video ingest. 

Figure 1 shows the top 10 systems for VDA maximum concurrent stream supported, where each stream equates to a single camera being recorded. 

Figure 1 SPECsfs2014_vdai top 10 maximum number of data streams supported

In Figure 1, the Quantum StorNext 7.0.1 came in as our new #1 with support for 7450 concurrent data streams. The StorNext system used all NVMe SSD storage with 14 (StorNext) client nodes, 1 metadata node and 10 (F-1000) storage nodes.

Each of the 10 F-1000 storage node had 10 15.4TB NVMe SSDs (total of 100 SSDs) and used RDMA storage access to the StorNext client nodes. They were all connected over a dual 100GbE network fabric and each StorNext node and F-1000 node had 2 ports to this network. 

StorNext used a separate, dual 1GbE management and metadata network to communicate with the 1 StorNext metadata controller node. 

In comparison, the DDN SFA14KXE (#2 in the chart above) was a disk-only system but it had  more disks  (454) and much more capacity (3600TiB) than the DATACOM storage. The WekaIO (#1 above ) was an all NVMe SSD, clustered storage system. 

Next, we show the top 10 Overall Response Time (ORT) for VDA submissions in Figure 2. 

Figure 2 SPEC sfs2014_vda top 10 VDA workload overall response times (ORT)

In Figure 2, the Quantum StorNext system came in as our new #1 with a 0.9msec ORT. This is the first VDA submission with an ORT under 1msec. Recall that ORT is an average across the whole benchmark series (from lowest to max stream counts). The Quantum 0.9msec response time was reached at ~5960 camera streams.

There’s a definite step function in ORT with the AFA solutions (#1-5) vs. the disk-only (#6) and hybrid SSD-disk (#7-10) systems. Moreover, the top 2 (Quantum and WekaIO) used all NVMe storage. It’s obvious why all NVMe and AFA storage would have the best ORT

In Figure 3,  we show the top 10 systems in max MB/Sec throughput for the VDA workload

Figure 3 SPEC sfs2014_vda top 10 VDA MB/Sec systemss

In Figure 3, we the new #1 in peak throughput is the Quantum StorNext with 34.4GB/Sec of throughput. This chart mimics the rankings of Figure 1 above, which implies that each stream is generated ~4.6MB of data per second, which corresponds to the SpecSFS, ~36Gbps specification for each stream

With the heavy write activity present in VDA camera streams, it’s somewhat surprising to see any all-flash storage doing well here. But all the AFA submissions had 1TB or more of memory in their submissions, so plenty of memory data and metadata caching was going on to achieve these throughput numbers. 

The Quantum F-1000 don’t use write caching. Each F-1000 had 64GB (640GB in total) of (read cache) memory, each client node had 192GB of memory for a total of 2688GB and the 1 metadata management node had 64GB of (meta-data cache) memory. The #1 StorNext had a total 3.4TB of memory across all their metadata, client and storage nodes and the #2 WekaIO submission had 11.7 TB of memory across their 23 nodes.

Significance

It’s always great to see new SPEC sfs2014submissions, even if there are only one of them. It gives us a chance to see and compare a set of storage systems under a defined load 

Quantum StorNext has been around for quite a while now (at version 7.0) and seems to be maturing nicely. With support for all NVMe/RDMA storage, StorNext is current with anything on the market today

We are still trying to determine the best way to report SPEC sfs2014 results and as such, we may experiment with variants of the above charts. Any ideas on other metrics of interest to report, please do let us know. Furthermore, suggestions on how to improve any of our performance analyses are always welcomed. 

This report was sent out to subscribers as part our free, monthly Storage Intelligence e-newsletter. If you are interested in receiving future storage performance analyses along with recent product announcement summaries, please use the QR code (below right) to sign up for your own copy.

[This storage performance was originally sent out to our newsletter subscribers in February 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.