[j-nsp] Juniper MX104

[Saku Ytti]

On (2013-11-12 20:25 -0500), Eric Van Tol wrote:

> One thing to keep in mind about these boxes is that, like the MX5/10/40/80, the built-in 10G ports do not do hierarchical QoS (per-unit scheduling).  I'm confused as to why this is, considering they are Trio-based routers, but I digress.  I personally don't think that the astronomical cost to enable the 10G ports on all the low-end MX routers is worth it, considering they can't even do per-unit scheduling.

It is annoying. As far as I know there is absolutely no technical reason why
the chassis ports couldn't use QX.

Only reason I can think of why they did that was to avoid grossly
oversubscribing QX chip. QX was dimensioned for MPC use-case, where you have
40G WAN and 40G fabric, so 40G was absolute maximum you'd need.
With MX80 and MX104, ports sit where fabric should be, so you have doubled
your demand for QX capacity.

QX performance is somewhere between 20Gbps (small packets) to 38Gbps (large
packets). And as far as I know this gets halved if you enable ingress and
egress shaping, so you might be looking as poor performance as 10Gbps.  So
maybe some PM thought it would be giving too much rope to customers to allow
enabling it on 8x10GE ports?

-- 
  ++ytti