Unlike our colleagues, we do not use NBench for explorations of that kind. Since modern hard drives come with as much as 16 megabytes of cache memory, we need a means to force as big an amount of data through the drive’s cache as to minimize the measurement error due to deferred writing. NBench would be always reporting a write speed higher than the linear read speed because the drive reports an end of a write operation before the tail of the file in the cache is actually written on the platter.
So, we use IOMeter instead and run it as long as to exceed the drive’s cache buffer in two times at least. We create four simultaneously running threads (an individual worker , each with its own portion of disk space to operate in, is responsible for each of the threads). The disk is accessed by 64KB blocks and the outstanding requests queue depth is steadily changing from 1 to 8.
The diagrams below show the results for the queue length of 1 as representative of the typical load on the disk subsystem of a desktop computer.
The case with one thread is in fact synonymous with linear reading from the disk. It’s no wonder then that the two WD740GD have identical results and the WD1500AHFD is much faster than them.
It’s different when there are two threads to be processed. The WD1500AHFD is still in the lead while the two WD740GD models have split up. The “wild” WD740GD-FLC0 is suddenly much slower than its brother. This is the consequence of its optimization for random operations and of its reluctant look-ahead reading.
I usually take the results of this test in by the ear. If the drive starts to click with its heads running between two work zones after switching from one thread to two, its speed is going to be low. But if the drive is almost silent, it is going to be fast in the test…
Here, the WD740GD-FLC0 drive didn’t believe that we’d continue read operations in either of the work zones and was receiving each read command without taking the statistics of previous requests into account.
The WD740GD-FLA1 is the fastest drive in the test at processing three threads; the WD740GD-FLC0 has accelerated and caught up with the WD1500AHFD.
The WD1500AHFD is unrivalled in the writing test. The WD740GD are delivering the same performance when processing one and two threads, but then the WD740GD-FLC0 suddenly leaves its opponent far behind. Or the WD740GD-FLA1 can be said to have fallen far behind its brother.
It’s hard to tell why the WD1500AHFD is superior in this test. It may be its larger cache or better firmware algorithms, but it is clearly faster than the previous-generation Raptors.