Switch recommendations? (fwd)

[Petr Ladislav Kodym]

		Hi,

Thanks for your very comprehensive answer. There might be a pitfall,
however.

>However, you're right -- the HP example alone indicates that I should
>stop worrying about what kind of switching technology a given switch is
>based on and focus on just the latency issue itself, since that is what
>matters.  Here is the corrected "rule":
>
>Switches that advertise latency in the 10 microsecond range or less are
>"good" (or at any rate "better":-) for fine grain packet traffic,
>whatever the mechanism they use for switching (which might be some
>proprietary mix of SnF and CT and black magic for all I really know
>about the internal hardware of a switch -- I just plug them in and they
>work, on a good day;-).  They also cost "more" (per port) and often
>come with other possibly desireable features, e.g. manageability, VLAN
>capabilities and so forth.

There are two ways of measuring the switch latency. HP reports latency of their
S&F switches on a LIFO basis (standard way of latency measurement for S&F
switches). LIFO latency is the time elapsed between the end of the last
bit of the packet going into the switch to the beginning of the first bit of that
packet emerging from the switch. Cut-through switch latency is however measured
on a FIFO basis --- time measured from the beginning of the first bit of the
packet going in the switch and beginning of the first bit of the outbound packet
coming out out of the switch. To convert LIFO figure into FIFO, one must add the
time spent on the wire:

  FIFO = LIFO+(packet length(bytes)*0.08 usec)

That doesn't make much difference for very short packets, but it adds 80 usec
for packets 1kB long. This could make a lot of difference if you are trying to go
user-level and bypass the TCP/IP stack overhead.

On the enclosed postscript figure are plotted our measurements of the TCP
round-trip times for different packet sizes using different types of
interconnections (3Com 3300 switch, SMC EZ1016DT switch and crossover cable).

All our machines were equipped with Intel EtherExpress Pro 100 NICs and they
were running the Linux kernel version 2.2.14. The line labeled as Uniproc Xover
shows the round-trip time between two single processor Athlon 500 Mhz machines
via cross-over cable. The line denoted as Uniproc SMC depicts the round-trip
time between the same pair of machines via SMC EZ1016DT switch. You can clearly
see the impact of the Store&Forward algorithm on the latency, even when using
high-overhead TCP/IP protocol stack. SMP Xover labels results of measurements
between two dual Celeron 466Mhz machines via cross over cable.  The initial
latency for 64B packets is almost twice higher than latency for uniprocessor
machines (why?). The last two plots denoted as SMP SMC and SMP 3Com displays
round-trip times between the two above described SMP machines over SMC EZ1016DT
switch and 3com SuperStack II 3300 switch correspondingly. Although the price of
these switches differs significantly, the performance from the point of view of
latency is exactly the same. SMC EZ1016DT switch has 16 ports, while 3Com 3300
comes in 12 or 24 ports version. However, SMC produces EZ1024 switch, equipped
with 24 ports as well. It's listed backplane bandwidth is 2.4Gbps, therefore it
should be capable of full wire speed on all 24 ports simultaneously. According
www.pricewatch.com this switch could be purchased by the beginning of February
2000 for about USD 520. According the same source, the street price for 3Com
SuperStack II was USD 1350 at the same time.


Hope that it was interesting for someone.



		 Petr