Victor Yodaiken wrote: > On Tue, Nov 27, 2001 at 11:49:46AM -0500, Stefan Doehla wrote: > >>Wasn't one of the goals of the whole APIC architecture reducing latencies??? >> > > No. Consider interrupt path in a SMP Intel IA32 board > assert-> I/O Apic -arbitration -> serial interrupt buss > | > processor apic > > You have a gain because you are not using the Pre-Cambrian era off chip > PIC on the standard UP Intel motherboard, but you have a loss because > of the arbitration, serial bus, and 2 levels of interrupt controller. > So it's kind of a wash.
This is the same I thought (but didn't know if it's true) as there are so many stages for the interrupts before they arrive at the right CPU. If one reads Microsoft papers or Novell developer guidelines - only use the APIC model ... Anyway, the APIC thing runs with a higher clock and should be programmable for smp affinity (I see there,s some code in RTL for this - I just don't understand all the details ...). Does RTL update the APIC register with the priorities of the tasks? Is it therefore possible to dedicate to one cpu only one interrupt when there's a realtime thread running on this cpu? Annotation: IRQ latency is 5-8 us with a non-SMP kernel. > We also believe that Linux performance is declining under the weight of > all the SMP "scaling" going on in 2.2+ and think this will be an issue > in 2.5 - but 2.5 is another headache. so is 2.4.x with SMP not 'realtime'-compatible? My tests here show that a 2.4.x kernel is not really realtime-stable. A 2.2.19 kernel is relative okay, but 2.4.4 SMP is good for milliseconds, not microseconds ... I wouldn't bother of latency if there is jitter in the range of 5-10us. The deadline for one execution here is 40 us (minus 20 us for I/O operations). RTL3.0, kernel 2.2.18, Redhat 6.2 was able to give us guaranteed 50-60us on a 1Ghz box (non-SMP). RTL 3.1 SMP with 2.2.19 provides 35-40 us but jitter (see my XFree4.x.x. mail). RTL 3.1 SMP with 2.4.4 is a candidate for the dustbin (fast and jitter like crazy). Am I running at the limit? The maximum speed later should be 20-25kHz - a watchdog will check if we are realtime ... Thanks for any comment or help Stefan -- [rtl] --- To unsubscribe: echo "unsubscribe rtl" | mail [EMAIL PROTECTED] OR echo "unsubscribe rtl <Your_email>" | mail [EMAIL PROTECTED] -- For more information on Real-Time Linux see: http://www.rtlinux.org/
