Paisit Blog

I have been working with storage performance for many years. Starting since 1999 when I have chance to conduct benchmark for application (ArcInfo) on Solaris. This ISV will choose Sun platform if performance on Sun platform (Server/Storage) can up to the mark. While we have quite number of tools to measure/analysis server performance, however, I found we are still lacking tools for storage. So that's starting point to keep me interested on how we can measure / analysis storage performance then lead to how we can improve performance of storage and eventually overall performance of customer application. 

Today, to solve server performance tend to be easier than storage since CPU power is increasing a lot of more and memory is cheaper and cheaper while even storage got cheaper (cost/GB) as well but storage have physic limit for performance (mechanical constraint - rotation per minute, still in milliseconds response time). So the gap is widen more and more (CPU go super fast, HDD improve a little).

In order to improve storage performance, we need to look into many aspects. For example, to sizing proper storage resource, it's depend on at least 3 factors (cost, performance, availability). The obvious example is RAID (0,1,5,6). We will get only 2 from 3. For instance, we can get RAID0 to provide us best performance and cost (no overhead i..e., parity) but we can not get availability. While on the server, availability can be separated from equation for sizing, hence, less complication. In addition, we need to look into access pattern from application perspective as well as what kind of storage technology to help. Today we have many storage technologies to choose such as tier storage (ILM), disk technology, controller technology, new storage protocol,etc. All can make storage performance analysis much more difficult.

When I start to study on this area, I have known a few people who inspired me. Some of them are still in Sun, some don't. Such as Brian Wong, Dave Fisk, Henk Vandenbergh, Allen Yen and etc. I met them in conferences and of course in the email alias where we have unlimited discussion. Several years back, Dave Fisk is the pioneer on this area and explain people that storage performance can be explained by queuing theory instead of treating it as the art. He shown us how "iostat" in Solaris OS work and relate to theory. It's my first time to see that storage performance can be explained by mathematics rather than treat it as the art. After he left Sun, we have Storage Performance Group - Steven Johnson, Amanda Hudson, Henk Vandenbergh and others who continue the idea/work until today. I have seen the way we deal with storage performance getting better. From command line of data collection/analysis which is very hard to understand by others, now we have fantastic GUI to work with. That make us much easier to deal with storage performance.