 |
Improve Linux performance
By Cameron Laird
2004-04-08
Reader Rating: 
Disk drives deserve doubt
A final tip for the performance-hungry is to be suspicious of disk drives. Mass storage has made amazing strides in density and reliability over the last few decades, but what I've seen lately is that at least some manufacturers appear to be stretching the quality of their operations too thin. Results include some units that fail outright and far more that simply give variable results.
Unless you measure carefully, you don't know the true average access time or throughput of your disk subsystems. This can be particularly frustrating when mass storage is configured as RAID, SAN, or other modern technologies. A particular unit might fail frequently, but the only consequence apparent at the application level is spooky variations in over-all performance. It can easily happen, for example, that most of the elapsed time of a particular program goes to error-correction within a RAID unit that has effectively lost one entire spindle.
What can you do about this? Plenty of remedies are possible, but they're poorly documented; I know them largely as the folklore that system administrators pass on in person. Here are a few highlights:
* Buy equipment you trust. Mass storage is a domain where you can pay for dependable products rather than accepting lowest bids.
* Don't push the envelope. Let others shake down SANs, the latest-generation SCSI, gigaether storage, and other offerings.
* Look for measurability in the products you buy. Find disk drives that monitor their own performance and circumstances, including temperature. Keep records so you know what to expect from your equipment. If your projects are big enough, you will run into disk mysteries. You'll be far better off with a documented baseline than trying to track down a failure after it has begun to happen.
Some performance fixes take only minutes to conceive or even implement; others require investments of days of research before they pay off. To track down the true limits on your applications' speed, you need to be prepared to tackle both the hard and the easy problems of performance.
First published by IBM developerWorks
If you found this article interesting, you may want to read these as well:
|
| |
|
|
|
