Elad Lahav wrote: > I am observing some strange numbers in /proc/stat while running a simple > micro-benchmark, which transmits UDP packets. The machine is a dual Xeon > with HyperThreading, which gives 4 logical processors, and 4 Gigabit > NICs, which I am trying to saturate. > > In the following experiment, I have pinned the transmitting process to > logical processors 2 and 3 (i.e., to physical processor #1), and the > network interrupts to logical processors 0 and 1 (two IRQs to each). > > Below is a snippet from mpstat, which reads /proc/stat. It is clear that > the affinity was set as expected. What I find strange is that logical > processor 2 seems to be handling a lot of soft IRQs, while LP 1's > numbers are close to 0. If I understand the kernel code correctly, then > soft IRQs are local, which means that only logical processors 0 and 1 > should be handling network soft IRQs. Moreover, IRQ numbers are at 0 > (which perhaps can be explained by good hard interrupt handling, which > does very little). > > In account_system_time(), the softirq field of the statistics is updated > if the SOFTIRQ_MASK flag is set, but I could not find where it was being > turned on (I expected it to be in do_softirq()). > > The options I can think of are: > 1. The statistics accounting is inaccurate, or even buggy > 2. Some other soft IRQ is executing on logical processor 2 > > I find option 2 unlikely, given that the machine does nothing else but > running the network-intensive benchmark. > > Ideas? > > Thanks, > Elad > > 03:32:02 PM CPU %user %nice %sys %iowait %irq %soft %steal > %idle intr/s > 03:32:04 PM all 1.35 0.00 16.95 0.00 0.00 6.14 0.00 > 75.55 8007.00 > 03:32:04 PM 0 0.00 0.00 0.00 0.00 0.00 0.50 0.00 > 99.50 4001.50 > 03:32:04 PM 1 0.00 0.00 0.50 0.00 0.00 0.50 0.00 > 99.01 4003.00 > 03:32:04 PM 2 5.56 0.00 63.89 0.00 0.00 22.69 0.00 > 7.87 0.50 > 03:32:04 PM 3 0.00 0.00 0.00 0.00 0.00 0.00 0.00 > 100.00 2.00 > 03:32:04 PM 4 0.00 0.00 0.00 0.00 0.00 0.00 0.00 > 0.00 0.00 > > 03:32:04 PM CPU %user %nice %sys %iowait %irq %soft %steal > %idle intr/s > 03:32:06 PM all 1.74 0.00 18.16 0.00 0.00 10.95 0.00 > 69.15 8047.24 > 03:32:06 PM 0 0.00 0.00 0.00 0.00 0.00 24.62 0.00 > 75.38 4022.11 > 03:32:06 PM 1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 > 100.00 4022.61 > 03:32:06 PM 2 1.51 0.00 13.07 0.00 0.00 4.02 0.00 > 81.41 1.01 > 03:32:06 PM 3 4.93 0.00 58.62 0.00 0.00 14.78 0.00 > 21.67 2.01 > 03:32:06 PM 4 0.00 0.00 0.00 0.00 0.00 0.00 0.00 > 0.00 0.00 > > 03:32:06 PM CPU %user %nice %sys %iowait %irq %soft %steal > %idle intr/s > 03:32:08 PM all 1.50 0.00 18.02 0.00 0.13 10.76 0.00 > 69.59 8007.00 > 03:32:08 PM 0 0.00 0.00 0.00 0.00 0.00 20.00 0.00 > 80.00 4001.00 > 03:32:08 PM 1 0.00 0.00 0.00 0.00 0.50 3.98 0.00 > 95.52 4003.00 > 03:32:08 PM 2 3.47 0.00 36.14 0.00 0.00 9.90 0.00 > 50.50 1.00 > 03:32:08 PM 3 3.03 0.00 35.86 0.00 0.00 9.60 0.00 > 51.52 2.00 > 03:32:08 PM 4 0.00 0.00 0.00 0.00 0.00 0.00 0.00 > 0.00 0.00 > I don't think I can really help you here, but I might as well note that I also had problems with /proc/stat. I have my questions regarding its accuracy: http://mail.nl.linux.org/kernelnewbies/2008-07/msg00196.html I should mention that when I increased the sleeping time my errors got relatively smaller. That is, if with 1 second of sleep I got values ranging from 99 to 500, with 10 seconds of sleep I got values ranging from 1000 to 1500. -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
