On 2019-09-02 17:16, Saku Ytti wrote:
On Mon, 2 Sep 2019 at 16:26, Denys Fedoryshchenko
<nuclear...@nuclearcat.com> wrote:
or some QFX, for example, Broadcom Tomahawk 32x100G switches only do
line-rate with >= 250B packets according to datasheets.
Only is peculiar term here. 100Gbps is 148Mpps, give or take 100PPM,
at 250B it's still some 50Mpps. Times 32 that's 1600Mpps, or 1.6Gpps.
Only implies it's modest compared to some other solution, what is that
solution? XEON doing ~nothing (not proper lookup even) is some couple
hundred Mpps, far cry from 1.6Gpps with ACL, QoS and L3 lookup.
I don't care about wire rate on chip with lot of ports, because
statistics. 250B average size on 32x100GE on a chip is fine to me.
250B average size on 32x100GE with 32 chips, would be horrifying.
I'm not saying XEON does not have application, I'm just saying XEON is
bps and pps expensive chip compared to almost anything out there,
however there are some application with very deep touch where it is
marketable.
Btw. technically Tomahawk and Trio are very different, Trio has tens
or hundreds of cores executing software, cores happen to have domain
specific instruction set, but still software box with lot of cores.
Tomahawk is pipeline box, having domain specific hardware and largely
not running a software (but all pipelines today are somewhat
programmable anyhow). On Trio you are mostly just time limited on what
you can do, on Tomahawk you have physical hardware restrictions on
what you can do.
Of course, they are much stronger (and cheaper in $/bps or $/pps) when
it comes to L2/L3 lookup, basic stateless filters, simple QoS.
But can Trio perform stateful firewall filtering for millions of flows+
lot of mpps that Xeon easily handle? Thats the case of recent DDoS
attacks.