Mag Gam put forth on 11/28/2010 7:10 AM: > Stan, > > thanks for the response. > > To my understanding, CONFIG_HZ is a kernel time option. Has that > changed? I can certainly rebuild the kernel. How can I check via /proc > what my HZ is currently set at? Is there a tool to determine this for > me?
I don't believe you can find this in /proc or /sys. On one of my servers, "grep CONFIG_HZ /boot/config-2.6.34.1" gives the following: # CONFIG_HZ_100 is not set CONFIG_HZ_250=y # CONFIG_HZ_300 is not set # CONFIG_HZ_1000 is not set CONFIG_HZ=250 I roll my own kernels from vanilla source and I use some options likely different than the Debian way of kernel building. I manually set CONFIG_HZ=250 on my servers these days as I don't run latency sensitive applications. As I stated, CONFIG_HZ sets the frequency of the kernel timer. This dictates, roughly, how many times per second the kernel interrupts itself to do housekeeping, such as process/thread scheduling, etc. The higher this value, in general, the more responsive the system is, yielding lower latency for applications. I learned this lesson long ago hosting multiple Half Life game servers, of all things. With a default Debian kernel, at that time, the kernel timer was 250Hz, common for "server" kernels where throughput is more important that application or network packet latency. Increasing the kernel timer to 1000Hz dropped game packet response times by a factor of 3, from say 75ms to 25ms. This completely changed the character of game play, eliminated lag, turned a lot of shot misses into hits, etc. I doubt one could find a more latency sensitive application than a first person shooter game server. Alas I've been away from that for a long time. :( > Removing tasks from cron has helped! We had some weird random tasks > starting up at production hours which causes interrupts. This is a > notoriously underestimated tip. When building HPC clusters from COTS boxes and off the shelf Linux distros, the very fist thing done to optimize MPI application performance is to remove all cron jobs and anything else that might fire off a process at any random time. Many/most distros create several default cron jobs, the most common being log rotation. In fact, on many HPC clusters, the OPs completely disable cron altogether. MPI applications are extremely latency sensitive, especially when using GigE as the interconnect instead of something like Infiniband, which has an inherent latency advantage on the order of 30:1 or more. -- Stan -- To UNSUBSCRIBE, email to debian-user-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org Archive: http://lists.debian.org/4cf2efd5.1050...@hardwarefreak.com
