cks@hawkwind.utcs.toronto.edu (Chris Siebenmann) (08/15/90)
The colour DECStation 5000 I just upgraded to Ultrix 4.0 is now getting about 280 interrupts a second even when it's totally idle. Quite frankly, I'd like the 10% of the CPU that they're eating up back. Unfortunately I have no idea what is causing them; they happen even when I boot the machine single-user. Has anyone else seen this behavior? Has anyone managed to work out why it happens and/or how to stop it? Thanks in advance. [*Sigh. New OS versions. I won't even mention how the multiuser boot sequence seems to start out on Wed Dec 31 19:00:02 EST 1969 about half the time.] -- "You don't *run* programs on Ultrix." - Mark Moraes "Right, you chase them." - Rayan Zachariassen cks@hawkwind.utcs.toronto.edu ...!{utgpu,utzoo,watmath}!utgpu!cks
mogul@wrl.dec.com (Jeffrey Mogul) (08/16/90)
In article <1990Aug15.024310.8785@jarvis.csri.toronto.edu> cks@hawkwind.utcs.toronto.edu (Chris Siebenmann) writes: > The colour DECStation 5000 I just upgraded to Ultrix 4.0 is now >getting about 280 interrupts a second even when it's totally idle. >Quite frankly, I'd like the 10% of the CPU that they're eating up >back. Unfortunately I have no idea what is causing them; they happen >even when I boot the machine single-user. > > Has anyone else seen this behavior? Has anyone managed to work out >why it happens and/or how to stop it? DECstations run with HZ = 256; i.e., the clock interrupt happens 256 times a second. The other 24 interrupts per second (or so) are probably related to I/O (such as file system "sync" activity, Ethernet broadcasts, etc.). It appears as if the way that vmstat and the kernel collaborate to decide how many interrupts/second are happening changed between 3.1D and 4.0, for MIPS (but not for Vax) systems. I.e., there's some code in vmstat.c to subtract HZ from the interrupt count for the Vax but not for the MIPS, and I suspect that someone changed the MIPS kernel to act like the Vax kernel but didn't change vmstat.c. Since the manual page for vmstat says "in" means "non clock" interrupts, I suppose this is a bug ... in vmstat, not the kernel. I seriously doubt that the clock interrupts are eating 10% of the CPU; not only does my DS5000 show about 98%-99% idle time (in "vmstat") but I've hacked my DS3100 (a slower machine) to run with HZ = 4096. It shows about 4120 interrupts/sec and also shows 98%-99% idle. So, I don't think the clock-interrupt handling consumes that much of the CPU. (I suppose "vmstat" might not be reporting the idle time accurately; how did you arrive at the 10% figure?) It is possible that your "extra" 24 interrupts/sec (which is about what I see, too) are caused by things that require your CPU to do something nasty (if each "wasted" 4 mSec, the total would be about 10% of the CPU). -Jeff