December 12th, 2001 by eugene
Capacity's Effect on Server Performance
In a meeting with Maxtor at Las Vegas' Comdex, we had an opportunity to pop the ever-present 15k question to Bill Clemmey, Senior Marketing Manager of Server Products. Bill's response, as always, was that while Maxtor was evaluating all options and had a team working on a potential "Atlas 15k," their research indicated that the market wasn't prepared to accept 15k RPM product in significant quantities due to cost and power concerns. In the mean time, he argued, partitioning a 73 GB Atlas 10k III drive to utilize just the outer 36 gigs would address those who required the utmost in performance.
A quick price check at HyperMicro.com places the 73 GB Atlas 10k III at $745. The Seagate Cheetah X15-36LP weighs in at $549. Thus, utilizing a 73 gig Atlas for 36 GB of storage is decidedly more costly when it comes to $/GB ratios.
Yet how would it perform? "Short-stroking" a drive through partitioning confines the drive's actuator into a narrower band of cylinders. As a result, the possibility of true full-strokes and other huge movements are eliminated. In this particular case, we're going to pit a 73 GB 10k RPM drive against a 36 GB 15k unit. Partitioning the former to utilize its outer 36 gigs of storage restricts actuator movement to less than half the platter's full breadth. It may very well eliminate the seek and latency advantages enjoyed by 15k RPM drives.
Our server benchmark, Intel's IOMeter, runs best when it has direct access to the drive. Partitioning and formatting a test unit unfortunately hinders IOMeter's precision. We bypassed this quandary by utilizing one of our newest tools, WinTrace32, to capture 10-minute trials of IOMeter's File Server and Web Server access patterns, each at 1, 4, 16, and 64 IOs, running on the Cheetah X15-36LP. Within these 80 minutes of activity, 1,184,172 requests were made to the Cheetah. The innermost request was for sector 71,682,019, or 36.701 GB into the drive. Hence, when RankDisk plays back these traces, no requests occur beyond this point. This effectively allows the exact same disk access recorded on the Cheetah to be played back on a variety of disks. Any disk significantly larger than 36 GB thus enjoys "short stroke" advantages.
These traces were then replayed on today's 73-gig SCSI drives: Maxtor's Atlas 10k III, Seagate's Cheetah 73LP, and Fujitsu's MAN3735. Unfortunately, at the time of this review, the IBM Ultrastar 73LZX is nowhere to be found. For control purposes, we also tossed in two 36 GB 15k RPM drives, IBM's Ultrastar 36Z15 and Fujitsu's MAM3367. Finally, we replayed the traces on Western Digital's Caviar WD1200BB to see if the drive's monstrous capacity allows ATA units to compete with SCSI drives in the multi-user domain.
IOMeter Score - in IOs/sec, the score of the drive as measured by IOMeter itself. These scores currently reside in the StorageReview.com database. The SR Server DriveMarks are normalized averages of these scores. These scores represent IOMeter's access patterns running across the entire drive.
% of X15-36LP Score - the % of a given drive's IOMeter score contrasted against that of the Cheetah X15-36LP.
IPEAK Score - in IOs/sec, the score of the drive as measured by a RankDisk playback of IOMeter's File Server and Web Server access patterns recorded by WinTrace32. These scores represent IOMeter's access patterns restricted to running within the drive's first 36 GB.
% of X15-36LP Score - the % of a drive's given IPEAK score contrasted against that of the drive from which the access patterns were recorded, the Cheetah X15-36LP.
Gain - The difference of the drive's IPEAK Score % and its IOMeter Score %. The greater the difference, the greater the gain achieved by short-stroking the drive.
As one would expect, there is little difference between the IOMeter score and IPEAK scores of the 36 GB drives. These again demonstrate IPEAK's efficiency in recording and replaying the data fed to it, a topic more thoroughly covered in this article.
Depending on the exact pattern and queue depth, the 73 GB, 10,000 RPM SCSI drives exhibit gains between 12% and 17%. In the case of the category-leading Fujitsu MAN3735, this is enough to bring performance within 2% of that of the Cheetah X15-36LP. Maxtor's Atlas 10k III also fares well, achieving roughly 90% of the X15's performance.
The Western Digital Caviar WD1200BB, however, fares poorly. Despite the 120 GB advantage, its gain even at a queue depth of one is not significantly greater than the SCSI drives. These gains also decay more rapidly as queue depths increase, presumably due to the ATA drive's lack of drive-level command queuing. This goes to show that while ATA drives are very competitive on the desktop, their relatively high seek times and lack of command queuing severely hamper server performance.
The conclusion? While partial-partitioning a drive yields seek time benefits that impact bottom-line performance, the gains achieved through short-stroking a 10k RPM drive aren't enough to erase the latency advantages of 15k RPM operation. Combine this with the fact that, for 36 GB usage, 73 GB 10k drives are more costly than 15k disks and its clear that if you need 15k performance, you should get a drive with a 15k RPM spindle speed. The only advantage of the 10k drives is slightly cooler operation- our top-plate measurements place the Atlas 10k III, for example, 3 degrees Celsius cooler than the Cheetah X15-36LP.