Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

[RE] Overprovisioning (was Re: [RE] CE applications)



There are two clear problems with overprovisioning.

1) No matter what level you set your network/interconnect, a reasonable
number of applications will cause queuing somewhere in the system (transmit
buffers, receive buffers, network (switch/router) buffers). The behavior of
these queues *must* be defined or the customer will be unhappy.
   1a) For example, when I was at Apple they thought the 1Gbyte/sec speed of
the HyperTransport interconnect was so "over provisioned" that they did not
test some interesting cases, and ran into serious buffer overflow problems
that could only be solved by throttling some much slower speed interfaces.
(The G4/PCI-based PowerBook that I'm using has better FW800 performance than
the G5/HyperTransport PowerMac that has so much better internal data rates.)
   1b) Another example is the fun that the enterprise people are having now
with queue overflows and lost performance in 1G and 10G SANs and cluster
connections. The bandwidth is *clearly* adequate, but the traffic models
cause all kinds of queuing problems.

2) Overprovisioning costs $. 1G is *a lot* more expensive that 100M. I can
*easily* integrate the 100M PHY into a larger part ... The cost of
integrating the 1G PHY into CE-type LSI is significant. ... And, yes, it
will get to a reasonable level in the not too distant future, but by that
time the uses of 1G bandwidths will be high enough that the queuing problems
in (1) above will show their ugly heads, and you will need 10G to avoid them
... And we are back to the $ problem.

So ... I think we should try to aim towards something that will *always*
work and not count on the *probability* that the consumer won't do anything
to cause queuing problems. Let's fix the problem and not pretend it doesn't
exist.

On 9/1/04 10:05 AM, "Gross, Kevin" <kevin.gross@CIRRUS.COM> wrote:

> Excuse me for responding to my own post but I'd also like to point out the
> overprovisoning solution to this scenario.
>
> Current home-class servers are challenged to saturate even a 100Mbit link
> with intelligent data. Even if neither the server nor client bandwidth were
> a bottleneck, bandwidth delay product issues inherent in TCP would likely
> slow down your file transfer and leave plenty of room for the streams.
>
> At 1Gbit, the network is clearly overprovisioned and would work fine even
> without 802.1Q.

--
-----------------------------------------------------------
Michael D. Johas Teener - Mike@Teener.com PGP ID 0x3179D202
23 Acacia Way, Santa Cruz, CA 95062-1313
+1-831-247-9666, fax +1-831-480-5845
------------------- www.teener.com ------------------------