Re: [fw-wiz] Cisco ASA and FWSM

On May 4, 2007, at 6:42 AM, <nick.nauwelaerts@xxxxxxxxxxx> wrote:
Timo Schoeler wrote:
yeah, and for the ASA-5520 (e.g.) they share one single interrupt.
worst hardware design ever.

Why would that be? Sharing interrupts will result in less context
switches. Different interrupts will result in a context switch for
interrupt, while shared interrupts can handle multiple interrupts
can be
handled in the same context switch.

There's going to be a context switch for each interrupt, regardless
of whether they are shared or not. It's typically the case that you
disable receiving new interrupts while running the interrupt service
routine (ISR), which means that you aren't going to get or handle
multiple interrupts in a single context switch.

Some of the better designed systems use fine-grained mutex locking
for each instance of a device driver and/or have a "fast interrupt
handler" which returns as quickly as it can to minimize interrupt
service latency, but depends on having a separate kernel thread to
perform the bulk of the work which needed to be done and would
normally have to be done within a traditionally designed interrupt
handler before it would be allowed to return.

Whenever you have a shared interrupt, all of the associated ISRs need
to run until one of them recognizes that the interrupt came from the
associated device, which causes more work and adds latency to the
ISR. And even if you have devices which have a well-written ISR
which is SMP-safe and does the necessary fine-grained locking, if
another device sharing that interrupt is not SMP-safe, then the
system will still have to obtain "Giant"/"the big kernel lock"/etc
for that ISR to run, which causes significant slowdowns.

In existing SMP x86 hardware, you'll commonly see that when you have
something like a USB port sharing an IRQ line with a NIC, the result
will be significantly reduced network performance.

And running their drivers in polling mode instead of interrupt mode
would make this even matter less, which
would make quite some sense in the ASA's case.

polling has some significant advantages in that it it tries to work
on a "process to completion" model and was designed to service
multiple outstanding requests during a single context switch, but you
end up running the polling service routine very often-- typically
some significant fraction, like 50%, of scheduler ticks-- and you'll
generally want to be using a scheduler quantum of around 1ms or so,
which means you're doing perhaps 500 polling ISRs per second
regardless of load.

If the device was seeing more than 500 interrupts/sec, then polling
would typically improve efficiency, but if it was mostly idle, it's
still going to be using a lot of CPU if polling is being used. Newer
NICs have much bigger packet buffers and can use interrupt mitigation
techinques to delay firing the IRQ line so that they might accumulate
several packets before the ISR activates, and thus they gain some of
the advantages of polling without also gaining the disadvantage of
keeping the CPU busy even when the device is mostly idle.


firewall-wizards mailing list

Relevant Pages

  • Re: Forth as an operating system
    ... calling upon an ISR to service an interrupt. ... Transputer) implemented compromises between preemption and cooperation ... but that instruction could be a context switch ...
  • Re: I/O issues, iowait problems, 2.4 v 2.6
    ... > OK, the IO rates are obviously very poor, and the context switch rate is ... That interrupt and context switch numbers are also _way_ out of line: ... And you should see context switches on maybe a request basis (ie you might ...
  • Re: shared interrupts, isrs and int-controller
    ... The OR of the outputs of the FFs is the processor's interrupt ... Thus the ISR is automatically repeated until ... I suppose that you are seeing a hardware or design problem. ...
  • Re: [Ksummit-2008-discuss] Delayed interrupt work, thread pools
    ... kind of context those operate into). ... Interrupt threads would be kernel threads and kernel threads ... that I need to context switch the SPU to other contexts while we block. ... Interrupts coming from SPU page faults will be deferred to a thread from ...
  • Re: Call for performance evaluation: (fwd)
    ... > don't occur often enough for their context switch time to matter. ... TCP performance drop off if I interrupt less often than once every ... Then there's latency. ... latency over sockets to risk using bizzare sockets-offload protocols ...