Re: a joint letter on low latency and Linux

[Benno Senoner]

David Lang wrote:
> Also note that if there are specific drivers that violate the latency
> timing but the kernel itself is good, then there will be a case of "if you
> need <5ms latency don't use a NE2000 card" or IDE drive or whatever.
>
> I am not listing this as a problem as much as a reality that will arise
> and an aswer to the suggestion that if the kernel can't be made perfect
> it's not worth doing
>
> David Lang

yes this is an issue.

eg old IDE disks are unsuitable for low latency because you can't turn
DMA on (you might risk disk corruption) and latencies will never go down
below 10-40msecs.

But my line is the following:
if you want to do serious work avoid flawed hardware.
non-flawed HW is cheap these days.

Regarding NICs I don't know if old ones can hurt latencies.
A rule of thumb would be all hardware which does slow PIO, 
transfers over the ISA bus, have smalll fifos which require a very high
IRQ rate in order to get throughput (with the risk of losing data)
etc etc

But almost all OSes will face these problems and I don't care
about the people owning crappy HW and at the same time wanting
to achieve killer latencies.

I care about getting good performance out of HW which CAN achieve
this performance (from a HW POV)

Benno.