Re: Beating system usage down
On Thu, Jun 24, 1999 at 12:34:06PM -0700, Mike Smith wrote: Just for those that have been following the benchmarking thread, this is exactly the same symptom set that FreeBSD demonstrates when loaded by WebBench. The gotcha here is, again, the giant kernel lock. Rather than trying to do the Solaris thing of mutexing everything, why don't we go in the opposite direction, and configure a multi-processor box as a cluster that happens to have really fast communications? Probably not as easy as it sounds, particularly since it would involve writing a memory network device driver, and some boot code to partition the main memory, and probably an extra layer of interrupt handling code, to hand device interrupts around. Er, yuck. It's just that it sounds as though it would be simpler to start with a blank sheet and a clean reentrant scheduling scheme, and graft pieces of FreeBSD back on top, than it would be to add that sort of functionality onto an existing traditionally structured Unix. -- Andrew To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
Beating system usage down
Greetings, A machine that I hold very close under my wing has been very contently chugging along for the last few months with practically no idle processor. However, I noticed that the CPUs are spinning a lot of cycles in the system area. CPU states: 5.5% user, 0.0% nice, 88.9% system, 4.0% interrupt, 1.6% idle First, some background. The machine is a Dual P2-450 with 1GB of RAM. It runs apache, and currently handles 90 hits a second, with each of those hits spawning various CGIs (one per hit) that completes in under a second. My first theory was that the kernel was uselessly spinning in various record locks via fcntl(). However, as a test I removed all file locking from the various CGIs and noticed no change in the system usage. My second theory was the overhead with the SMP code. So, I removed it from the kernel and ran a single CPU box for a few minutes. The system usage went down to around 60%, but the system was noticeably slower. Any ideas? Regards, Dan To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message
Re: Beating system usage down
Just for those that have been following the benchmarking thread, this is exactly the same symptom set that FreeBSD demonstrates when loaded by WebBench. The gotcha here is, again, the giant kernel lock. Greetings, A machine that I hold very close under my wing has been very contently chugging along for the last few months with practically no idle processor. However, I noticed that the CPUs are spinning a lot of cycles in the system area. CPU states: 5.5% user, 0.0% nice, 88.9% system, 4.0% interrupt, 1.6% idle First, some background. The machine is a Dual P2-450 with 1GB of RAM. It runs apache, and currently handles 90 hits a second, with each of those hits spawning various CGIs (one per hit) that completes in under a second. My first theory was that the kernel was uselessly spinning in various record locks via fcntl(). However, as a test I removed all file locking from the various CGIs and noticed no change in the system usage. My second theory was the overhead with the SMP code. So, I removed it from the kernel and ran a single CPU box for a few minutes. The system usage went down to around 60%, but the system was noticeably slower. Any ideas? Regards, Dan To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message -- \\ The mind's the standard \\ Mike Smith \\ of the man. \\ msm...@freebsd.org \\-- Joseph Merrick \\ msm...@cdrom.com To Unsubscribe: send mail to majord...@freebsd.org with unsubscribe freebsd-hackers in the body of the message