24.11.2012 Views

Traffic Management for the Available Bit Rate (ABR) Service in ...

Traffic Management for the Available Bit Rate (ABR) Service in ...

Traffic Management for the Available Bit Rate (ABR) Service in ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

4. Large legacy switches have a problem <strong>in</strong> that <strong>the</strong>y are very expensive to replace,<br />

but provide few hooks <strong>for</strong> add<strong>in</strong>g new functionality. Counters, registers and<br />

p<strong>in</strong>-outs are always at a premium on a chip, and are rarely left unused. For<br />

example, one might decide to use a switch hook (a p<strong>in</strong> <strong>in</strong> <strong>the</strong> chip) to implement<br />

an improved switch scheme. However, if <strong>the</strong> scheme requires measurement of<br />

di erent quantities, it cannot be done due to <strong>the</strong> lack of hooks. For example, <strong>the</strong><br />

ERICA switch scheme requires <strong>the</strong> measurement ofquantities such as <strong>the</strong> <strong>in</strong>put<br />

rate and <strong>the</strong> number of active <strong>ABR</strong> sources. But, <strong>the</strong> only available metric <strong>in</strong> <strong>the</strong><br />

switch might be just <strong>the</strong> queue length, which doesnotallow <strong>the</strong> implementation<br />

of <strong>the</strong> algorithm to be retro tted on <strong>the</strong> switch. Simple algorithms which use<br />

just <strong>the</strong> queue metric need to be used <strong>for</strong> such cases.<br />

5. Some <strong>ABR</strong> features such as <strong>the</strong> Virtual Source/Virtual Dest<strong>in</strong>ation feature re-<br />

quire <strong>the</strong> implementation of per-VC queue<strong>in</strong>g. Recall that <strong>the</strong> requirement of<br />

per-VC queue<strong>in</strong>g was one of <strong>the</strong> key reasons why <strong>the</strong> credit-based framework<br />

was rejected. It might seemcontradictory to see per-VC queue<strong>in</strong>g implemented<br />

by all major vendors. However, note that <strong>the</strong> current switches typically support<br />

upto 128K VCs per port. When <strong>the</strong> number of VCs grows fur<strong>the</strong>r (millions of<br />

VCs), <strong>the</strong> account<strong>in</strong>g <strong>in</strong><strong>for</strong>mation required and <strong>the</strong> schedul<strong>in</strong>g overhead is ex-<br />

pected to become prohibitively expensive. In such cases, VCs will be aggregated<br />

<strong>in</strong>to classes and supported by afewthousand class queues.<br />

VS/VD, on <strong>the</strong> o<strong>the</strong>r hand, requires <strong>the</strong> implementation of <strong>the</strong> source end sys-<br />

tem rules, schedul<strong>in</strong>g of VC cells at a variable ACR, and <strong>the</strong> ma<strong>in</strong>ta<strong>in</strong>ence of<br />

a large amount of state per-VC. This has resulted <strong>in</strong> <strong>the</strong> VS/VD option to<br />

be implemented only <strong>in</strong> very large switches (like satellite switches) where <strong>the</strong><br />

370

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!