Performance in Intel IOMeter

Database Patterns

In the Database pattern the disk array is processing a stream of requests to read and write 8KB random-address data blocks. The ratio of read to write requests is changing from 0% to 100% (stepping 10%) throughout the test while the request queue depth varies from 1 to 256.

We will be discussing graphs and diagrams but you can view the data in tabled format using the following links:

• Database, RAID1+RAID10
• Database, RAID5+RAID6

Everything is normal at a queue depth of 1. The arrays show good scalability. The degraded RAID10 does not slow down and, together with the normal 8-disk RAID10, coincides with the 4-disk RAID0, just as the theory has it.

The checksum-based arrays are very good at a queue depth of 1, too. In our earlier tests with four WD Raptors with SATA interface we could observe a sudden performance hit at write operations, but the new generation of RAID controllers proves that a 4-disk RAID6 can be as fast as a single HDD thanks to the large amount of cache memory, low disk access time and fast architecture. The 4-disk RAID5 proves to be faster than the single disk at writing. The 8-disk arrays are far ahead.

The RAID6 slows down but a little when one of its disks fails but the failure of two disks leads to a performance hit. Having to constantly restore data from two checksums (writing to a rotated parity array involves reading operations), the controller gets much slower: the degraded array falls behind the single disk. The RAID5 suffers greatly even from losing one disk: its performance is worse than that of the RAID6 without one disk, yet not as bad as that of the RAID6 without two disks.

The controller’s performance grows up at a queue depth of 16 requests. The RAID10 arrays are close to the RAID0 consisting of half the number of disks at high percentages of writes. At high percentages of reads, they are close to the RAID0 arrays consisting of the same number of disks. The RAID10 are not faster than the RAID0 at reading, which means that the controller can read from both disks in a mirror but has no read optimization (it does not give the read request to what disk in the mirror can perform it faster).

The degraded RAID10 does not show up its degradedness at high percentages of writes, which is good, but loses its ground at high percentages of reads. Its performance does not slump to the level of the 4-disk arrays, meaning that the controller is reading from both disks in the non-defective mirrors.

The RAID5 and RAID6 arrays speed up, too. The RAID6 arrays are nearly all right whereas the RAID5 are far from ideal. The latter’s graphs fluctuate, indicating some flaws in the controller’s firmware, especially with respect to degraded RAID5.

We say that the RAID6 are nearly all right because of the 4-disk array’s performance: it is unexpectedly slower than the single disk at high percentages of writes. The 8-disk array survives the loss of one disk more or less acceptably but the loss of a second disk affects its performance greatly: the degraded RAID6 without two disks is ahead of the single HDD at reading but cannot compete with the latter at writing.

When the queue depth grows even longer, we have to take back the supposition we have expressed above: the controller can actually find the faster disk in a mirror. This is especially conspicuous if you compare the 4-disk arrays: the RAID10 is much faster than the RAID0 at pure reading. The degraded RAID10 is a disappointment. It copes with writing well but fails at reading. As a result, it is just slightly better than the 4-disk RAID0 through most loads.

It is interesting to compare these numbers to those of the 3ware 9690SA controller that has a processor with completely different architecture. The two products are overall equals: the Adaptec is somewhat more effective at writing whereas the 3ware is somewhat faster at high percentages of reads. Besides, the 3ware’s degraded RAID10 has a smaller performance hit.

There are no serious changes in the standings of this group except that the RAID5 arrays are now somewhat better at reading, especially the degraded array. As a result, we see an odd picture with the 8-disk arrays: a RAID6 is ahead of a RAID5 at pure reading! The degraded RAID6 behave in the same way: the array slows down without one disk and slumps below the level of the single HDD when two of its disks fail.

If we compare this to the results of the 3ware 9690SA controller, we can see that the latter is better with RAID6 under any loads and is faster with RAID5 at reading. The Adaptec’s RAID5 arrays are faster at high percentages of writes.