Quantum Fireball SE: SCSI vs ATA

The ATA Fireball SE bests the SCSI version by 13% in the Business Disk WinMark 98 and by 15% in the High-End Disk WinMark. Curiously the results for the ATA drive were also substantially better in the low-level tests in addition to the Disk WinMarks. Both drives supposedly have the same seek time and the same rotational speed (and thus latency) yet the ATA Fireball posted a 16.3ms average access time to the SCSI version's 18.0ms. The Disk/Read Transfer rates of the SCSI Fireball also lagged behind its ATA counterpart.

CPU utilization per megabyte transferred for both drives are virtually identical. There's a common perception floating around that ATA's CPU utilization even with busmaster drivers does not approach that of busmastering SCSI. CPU utilization scores of ATA drives when using the busmaster drivers including with Windows 95 OEM Service Release 2 indicate otherwise, however. SCSI does not have an advantage here.

ATA drives and controllers, on the other hand, do not offer a key SCSI feature: Command Queuing. This feature allows multiple requests to "queue" and remain pending on the SCSI subsystem as opposed to ATA's serial approach where one request must be finished before another is started. This parallel execution allows SCSI to reorder the requests into a sequence that minimizes head/actuator seek motions. Simply put, SCSI's advanced features should result in less degradation in multitasking and multithreaded environments.

We decided to turn to Adaptec's ThreadMark 2.0 Benchmark. Adaptec offers the benchmark as a more reliable indicator of overall disk performance in multitasking environments. ThreadMark is difficult to run consistently, however. To obtain repeatable results, the test machine's use of virtual memory must be disabled. Storage Review's testbed has 64 megs of RAM, enough to run most applications well, but -not- enough to guarantee stability without any virtual memory. With virtual memory enabled, ThreadMark results varied wildly, sometimes displaying sample deviations of over 100%. Disabling virtual memory contained the sample deviations to a much more usable 1%. The average of five trials is presented here.