Windows 2000 IDE I/O Performance
In the introduction, we showed how IDE disks are substantially cheaper than SCSI disks per GB of capacity and per disk arm. In these next sections, we look at IDE performance to see what these inexpensive drives are capable of.
Windows 2000 IDE Random I/O
Unbuffered 8KB requests randomly spread across the entire drive delivered approximately 100 IOs per second using 32-deep requests. At 5400 RPM, this is 1.1 IOs per revolution, the same number we saw from tests on our 10K RPM SCSI drives. As the request depth increases from 1 to 32, the driver, controller, and the disk drive are able to optimize the arm movement in order to service the requests more efficiently. One and two-deep requests show 66 IOs per second for reads and 92 IOs per second for writes – 37% lower on reads and 12% lower on writes than that of a single SCSI Quantum Atlas 10K drive. This is because at one-deep, the accesses are random with an average service time of 15ms for reads. At two-deep, as soon as the first request comes in, the drive begins to service the request. The second request arrives after the drive has been committed to servicing the first request, eliminating the possibility of reordering. The net result is read performance equal to one deep performance. Beyond two-deep read requests, the controller is able to optimize the order in which requests are fulfilled, and so the read IO rate gradually increases. Writes are buffered by WCE, hence 1-deep requests in the file system are potentially very deep in the controller. The controller can sort the random write requests, and so get shorter average seek times.
Random IO performance is intimately linked to drive latency. Each time the heads seek to a new track, they must wait for the right sectors to travel under the drive head. As we increase the number of RPMs, the drive latency decreases so the drive waits for a shorter period of time before the desired sectors travel under the drive head. Considering that a 5400RPM drive rotates 46% slower than a 10K RPM drive, it is easy to see the effects of rotation rate. Similarly, the Quantum Atlas 10K had an advertised seek time of 5.0ms while the Quantum Fireball’s advertised seek time is 9.5ms which is 48% slower. The Atlas 10K SCSI drive is about two and a half times more expensive than a Fireball IDE drive, so IDE turns out to be cheaper in terms of IOs per dollar. At $209, the Quantum Fireball costs $3.17 per KAPS (kilobyte access per second) while the Atlas 10K costs $534 or $5.09 per Kaps. The number of IOs per second scales linearly with the number of drives, so two and a half drives would give us 2.5*(66) IOs per second or 165 IOs per second. Compared with the 105 IOs per second achieved on an Atlas 10K, for the same amount of money, you get 57% more IOs per second than a single SCSI drive.
|
