Performance Management: Day 3
Nov. 26th, 2008 08:36 pmAfter yesterday's session on LVM performance, today naturally segued into a discussion of JFS tuning tips. Some of these were so obvious, such as not enabling dynamic compression, they merely needed a single comment. Some of the others, such as release behind buffering or IO pacing, were covered in greater depth.
After a quick lab session, setting up and defragging file systems, we went on to look at TCP/IP and some of the factors that govern its performance. There were some interesting comments about socket buffer tuning, while both Nagle's algorithm and sliding socket windows were covered clearly and concisely.
The lab following the exercise, however, was a less than comprehensive success. We first found ourselves unable to measure a meaningful round trip time between the different systems we were using — because they were LPARs, perhaps? — before going on to discover that this lack of latency did not mean that the network connection was particularly reliable if dismal performance and number of retransmits was anything to go by. In the end, we abandoned all hope and deferred the rest of the exercise until tomorrow morning.
After a quick lab session, setting up and defragging file systems, we went on to look at TCP/IP and some of the factors that govern its performance. There were some interesting comments about socket buffer tuning, while both Nagle's algorithm and sliding socket windows were covered clearly and concisely.
The lab following the exercise, however, was a less than comprehensive success. We first found ourselves unable to measure a meaningful round trip time between the different systems we were using — because they were LPARs, perhaps? — before going on to discover that this lack of latency did not mean that the network connection was particularly reliable if dismal performance and number of retransmits was anything to go by. In the end, we abandoned all hope and deferred the rest of the exercise until tomorrow morning.