RE: [10GBASE-T] latency
Bruce,
It would be
most helpful to have a survey of applications and latency requirements. It might
be easier if we can restrict this survey to the space of interest. In short, if
the question could be bounded (will do no worse than X, can do no better than
Y), then it would be far simpler for said experts to
respond.
We might
anticipate responses:
1. Y is
not good enough. Unless you can get to Y2 (< Y) my application won't
run.
2. X is good
enough. Make it as simple as possible.
3. I can run
between X & Y, prefer Y, but don't want to pay much over X for
it.
In order to
help set these boundaries, there are certain classes of problems that can be
moved to NUMA machines if the latency is sub-microsecond, process to
process.
jonathan
-----Original
Message-----
From: Bruce Tolley
[mailto:btolley@cisco.com]
Sent: Monday, February 23, 2004 11:17
AM
To: jonathan.thatcher@ieee.org;
'stds-802-3-10gbt'
Subject: RE: [10GBASE-T]
latency
Jonathan
Thanks
for the summary
I would argue that early 10GBASE_T switching products
should be sold to early adopters at National Labs and other R&D sites
building clusters. We need this community to come to the TF and state its
latency requirement in the 2006 timeframe and determine the
tradeoffs.
Bruce
At 10:54 AM 2/23/2004 -0800, Jonathan
Thatcher wrote:
There have
been numerous interesting and correct comments made. A subset of these apply
only in certain contexts. To that end, I will attempt to add some
context.
There is little question that lower latency increases the
market potential. There is little question that lower prices (read that less
complexity), and earlier time to market also increases the potential market.
The problem is that these fight against each other, and the optimization
point is not clear.
I presume that there are two principal application
spaces for 10GBASE-T in the near term: data center and enterprise (home and
school will probably have to wait a couple of years :-). If you want a
strict boundary between these two spaces, I can't provide it. So we will
have to deal with some ambiguity. In the enterprise, it is difficult to
argue that low latency is as critical as low price. The exception to this
would be low latency applications that want to be set up as a "grid
computer," which I will lump into the "data center" bucket.
The data
center, on the other hand, has instances where both low latency is required
(clustered computing) and higher latency is acceptable (most file serving).
From a parallel computing perspective, there are classes of problems
(applications) that range from low latency NUMA to those that are
"embarrassingly parallel (e.g. http://setiathome.ssl.berkeley.edu/)."
From the
perspective of the upcoming "Data Center Ethernet" (may not be the best
name) call for interest, the intent is to explore those means that can be
used to decrease latency in Ethernet networks. If one is to presume that
this should be a key application space for 10GBASE-T, then it would be
interesting to understand the trade-off between latency and complexity. It
may be the case, that even under the most complex scenario, that 10GBASE-T
latency is simply insufficient for entire classes of low latency
applications.
So, the question remains, what does the complexity
vs latency curve look like? I expect that it is something like the left side
of a bathtub curve (vertical axis is latency, horizontal axis is
complexity). What is the inflection point? What is the slope of the falling
portion of the curve? What is the asymptote?
jonathan
Bruce Tolley
Senior Manager, Emerging Technologies
Gigabit Systems Business Unit
Cisco Systems
170 West Tasman Drive
MS SJ B2
San Jose, CA 95134-1706
internet: btolley@cisco.com
ip phone: 408-526-4534
"Don't put your hiking boots in the oven unless you plan on eating
them."
Colin Fletcher, The Complete Walker