Performance in Intel IOMeter: DataBase Pattern

The Database is traditionally the first pattern in our test program. It helps us to evaluate a drive’s performance with a mixed stream of requests for reading/writing random-address 8KB data blocks. By changing the ratio of reads to writes we can check the drive’s ability to sort the requests out.

The results table is very big, so we divide the results of this pattern in groups, by manufacturer.

Hitachi

For better readability, we draw diagrams for different workloads (the number of requests in the queue).

You can see it in the above diagram that the drive with a 2MB cache buffer starts losing to the other models from 50% writes on. It comes as all HDDs from Hitachi are very unwilling to give the cache segments reserved for reads, for writes. A drive with a small cache buffer has overall fewer cache segments than an 8MB-buffered device. When a drive receives only write requests, the drive’s firmware uses the whole cache to store writes, and the random write speed increases suddenly, due to the more efficient work of deferred write algorithms.

The IC35L180AVV207-1 model from the older Deskstar 180GXP series loses to the members of the new Deskstar 7K250 family when there are many writes to be performed. As for a difference between the 7K250 disks with an 8MB buffer but different interfaces, there is none.

We increase now the load to 16 requests:

The drives all have the same speed in the random read mode, but split up in two groups as soon as there appear write requests. The first group is comprised of the 7K250 disks with an 8MB buffer – their speeds are actually much alike.

The remaining two, the IC35L180AVV207-1 and the HDS722516VLAT20, started out together, but the latter couldn’t keep the pace on due to its small cache. You should also note that the HDS722516VLAT20 suddenly accelerates at the 100% writes point, just it did under the 1-request load, and stands even with the other drives there.

The workload being the highest, the graphs practically remained the same as under the 16-requests load.