Notes About Testing
Without support for iSCSI, we can't run a lot of our favorite tests. So, we're implementing a new process for benchmarking network-attached storage. We've actually been planning this for a year but are just getting around to utilizing it. Going forward, you will see performance results with SMB and iSCSI, though this review limits us to SMB.
We chose two products to test against Drobo's B810n: the Netgear ReadyNAS 716 and the TS-809 Pro, an older appliance from QNAP. We tested the Drobo in two configurations, both using hard drives in RAID 6. Our baseline employs eight Seagate Enterprise NAS 6TB disks, while the second setup utilizes Drobo's auto-tiering function with three SanDisk Extreme Pro 240GB SSDs and five Seagate Enterprise NAS 6TB hard drives. You can use a single SSD for caching, but to fully benefit from auto-tiering, three SSDs are required.
Sequential Read Performance
Every system we've tested with a Marvell processor has displayed some type of abnormality in the test results. That isn't always a bad thing. One low-cost NAS we reviewed actually demonstrated better sequential write performance than anything else in its class. Most of the time, though, the odd behavior goes the other way and we uncover lower performance than competing Intel-based machines.
Indeed, the B810n encounters a dip in our sequential read test with 512KB blocks, and it happens with or without caching. When we enable the cache, Drobo's B810n does perform at the same levels as its competition with few outstanding commands.
Sequential Write Performance
The sequential write tests show more consistency, though we don't see much of a speed-up from auto-tiering. Then again, we usually don't expect a boost from flash in sequential workloads since the feature is tuned to accelerate random transfers of small blocks.
Sequential Mixed Performance
Mixed workloads are often overlooked in network storage testing. But if you run software directly from the NAS, a lot of files get updated and the system juggles data moving in both directions at the same time. Professional photographers have a long history of using Drobo, and many edit their work directly from networked storage. Because of that unique market, we think sequential mixed workloads are more prevalent than the random mixed workloads that take precedence on iSCSI-equipped appliances.
The B810n performs well here, and it fares even better with auto-tiering enabled. I suspect that Drobo plays off of its customer base and tunes the system for this type of workload.
Random Read Performance
Our random read and write tests make it pretty clear that Marvell's quad-core SoC isn't as powerful as competing solutions from Intel. Drobo's B810n performs poorly when we tax it with small blocks of data read and written randomly, particularly when it's only armed with mechanical storage. Of course, performance improves with a flash-based front end, as you'd expect. But the flash is still held back by low processor performance. When we add a solid-state cache to competing platforms, random read performance often increases beyond 2000 IOPS. The cache-enabled B810n only manages to match the other systems with an Intel SoC and disk drives (no flash-based cache).
Random Write Performance
Random write performance is very low when we only arm the B810n with hard drives. Again, flash helps, but not as much as it benefits the Intel-based NAS servers. Some systems, like Netgear's RN716X and many products from Asustor, use DRAM as a random write buffer. That's what we see from Netgear's entry.
One phenomenon we observed from the B810n was that random write performance held steady, even when we increased block size. QNAP demonstrates similar behavior, and that's why we chose it to appear in today's review. With most products, IOPS drop as the block size increases.
Random Mixed Performance
Random and mixed workloads stress the processor. Combining both yields a worst-case scenario for most networked storage devices. With auto-tiering, the B810n manages to outperform the other models in our charts. But without the aid of flash, performance drops to very low levels. We like to see numbers at or above what a single hard drive delivers attached directly to its host via SATA, but that's not the case with Drobo's B810n.
Standard Server Workloads
Our server workloads model popular tasks performed with network-attached storage. We don't see anything new in this set of tests that we didn't already observe on the previous page. Drobo's B810n suffers from poor performance when it's loaded with mechanical disks, but catches up when it's complemented by a flash-based cache. The latency over IOPS charts prove that this system is underpowered in heavy workloads.
Trace-Based SMB Workloads
When we lighten the workload to match desktop-oriented apps, the story changes. In many of these tests, our results are much better than the synthetic benchmarks had us expecting. That's good, since the tests in this section are more inline with what Drobo tuned its platform to handle. Therefore, we weigh these tests more heavily than the others.
Multi-Client Office Workloads
Drobo says its auto-tiering feature is a way to maintain performance when several clients access the NAS simultaneously. We find this to be true, though it's not effective if the workload increases beyond a certain point. Moreover, auto-tiering didn't pick up our Office workload, and that is troubling since the traces are recorded with real Microsoft Office applications running on another storage platform. This is the first system we've tested with an SSD-based cache that failed to detect and accelerate Office back to individual clients.