SSD and RAID - Benchmarks and Conclusion

SSDs In RAID Storage
By , Manuel Masiero

Adding more drives increases performance and capacity in equal measure our test of SSD in RAID 0. The same applies to both data throughput and I/O operations per second.

Test Setup

The five SSDs used in the test come from Samsung. They do not have any strong product name, carrying only “Samsung SLC SSD“ and the product tag MZ3S9100-XAB4. Each drive has a capacity of 100 GB, is based on SLC NAND flash, has a 3 Gb/s SATA interface, and support for TRIM and Native Command Queuing (NCQ). Samsung claims 260 MB/s read and 245 MB/s write rates.

With an idle energy consumption of 1.7 W, this SSD requires much more than consumer models. It reaches 3.7 W under load, which is still acceptable compared to enterprise-class hard drives spinning at 10 000 and 15 000 RPM. Conventional disks can reach up to four times that number. Operating temperature is rated at between -10°C and 60°C, which is fairly standard in the SSD space. Samsung also offers a 200 GB model, should you need more server storage capacity.

In order to demonstrate the scalability as clearly as possible, we first benchmarked one of these SSDs and then created a RAID 0 array of two drives. With each new benchmark run, we added an additional drive so that, in the end, we can compare five test configurations and their differences easily.

Of course, the fast (but unsafe) RAID 0 mode is only used in special cases, while RAID modes with built-in redundancy are typically preferred. For this test however, we deliberately chose RAID 0, since it represents a best case for the kind of scaling we can expect.

Although we're using a relatively fast test system, the configuration is not ideal to maximize I/O performance. Multi-socket platforms, faster processors, and one of the latest RAID controllers would help to maximize I/O performance numbers compared with the results we see. We also have to say that enterprise-class SSDs are not necessarily best at delivering maximum performance, but at maintaining performance. In this case, it means that other SSDs, especially consumer products, may appear faster on paper and according to basic benchmarks. However, when hammering them with intensive workloads, enterprise drives are better able to maintain expected performance levels at all times.

I/O Performance

The Iometer benchmark and its File Server, Web Server, and Workstation benchmarks illustrate how the system scales. While a single SSD reaches approximately 15 000 I/O operations per second, a RAID 0 array with two drives gets an additional 10 000 I/O operations. Adding more SSDs increases the total I/O capabilities by roughly the same number.

Each added SSD not only increases capacity, but I/O performance grows strongly and steadily as well. The test system does not seem to be nearing any upper limit of I/O performance, but scales with the same amount for each drive added. Using traditional magnetic hard drives, such performance is only reached by adding a really substantial number drives (in the double-digits). Or, to put it differently, where you once had to add several new servers or huge RAID arrays, buying another SSD might be enough in today's landscape.

The Database benchmark falls somewhat out of place, as a single SSD should achieve better performance. The trend still remains the same, though, with a five-drive RAID 0 array providing much higher I/O throughput.

 Iometer Streaming

Speed enthusiasts will get their money’s worth when it comes to sequential read rates. With five drives, the RAID array has peak data throughputs of 1000 MB/s (read) and 1122 MB/s (write). The latter is faster due to the RAID controller’s caching. More importantly, each added drive increases the performance by about the same value, scaling almost linearly.

4 KB Random Reads/Writes

When reading and writing randomly-selected 4 KB blocks, the results look like this. Adding more SSDs increases the data throughput and I/O operations per second. As noted earlier, the performance scaling is almost linear. Only at a queue depth of one does the number of added drives becomes unimportant.

PCMark Vantage

In contrast to Iometer and CrystalDiskMark, PC Mark Vantage shows significantly less differences when adding more SSDs. The RAID array is still better than a single drive, but the scaling is much worse. This is mostly due to the nature of the benchmark itself, however. It mainly focuses on consumer scenarios, and we chose to include it mostly for the sake of painting a bigger picture.

The difference is smallest in the gaming benchmarks, since the HDD data rate only plays a small part of gaming performance.

The greatest performance increase is seen in the Windows Media Center tests, as it measures the streaming capabilities of the system.

Conclusion

This test of SSDs in a RAID 0 array brings to light an interesting effect: adding more drives increases performance and capacity in equal measure. The same conclusion applies both to data throughput and I/O operations per second.

Performance scales almost linearly in our test system, each added drive offering roughly the same increase. When it comes to I/O performance, we did not run into any upper limits or decreasing scaleability when adding drives, so you should be able to extrapolate the results for your own scenario up to the bandwidth and latency limits of your platform architecture. Expanding capacity brings equal performance increases with SSD RAID, which is a first in this area.

Of course, the results cannot be perfectly applied to all RAID systems, but the test brings a key finding: an SSD-based RAID array benefits much more from adding more drives compared to traditional hard drives. The more flash drives involved, the greater their strengths become, especially the I/O performance that is so crucial in the business segment. While you probably won't see RAID 0 arrays like these used in many production servers (they'd be a little more plausible in a video workstation, perhaps), there are other performance-enhancing RAID modes you could use to augment the speed and reliability of a mission-critical system.

For enterprises, this provides a new criterion that should be considered when planning an SSD RAID system. In addition to the prices of flash drives, controllers, and the server platfrom, one also has to take into account the number of drives that are going to be used. However, the test results show that the strengths of SSD RAID can only be properly exploited in an appropriate environment, such as a high-speed data access-dependent Web or file server.

Comments