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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

2. Some switch schemes have a process<strong>in</strong>g requirement <strong>for</strong> both <strong>the</strong> <strong>for</strong>ward and<br />

backward go<strong>in</strong>g RM cell. This requires extra process<strong>in</strong>g at <strong>the</strong> switch. Ano<strong>the</strong>r<br />

related problem which arises <strong>in</strong> this case is that <strong>the</strong> switch scheme requires<br />

exchange of <strong>in</strong><strong>for</strong>mation from one port to ano<strong>the</strong>r (<strong>the</strong> <strong>in</strong>gress chip to <strong>the</strong> egress<br />

chip). The assumption made by <strong>the</strong> scheme is that <strong>the</strong> switch has a shared<br />

memory which is used <strong>for</strong> tables facilitat<strong>in</strong>g <strong>the</strong> <strong>in</strong><strong>for</strong>mation exchange. This<br />

assumption is based on <strong>the</strong> fact that early switches had <strong>the</strong> VC table (which<br />

maps a cell of a VC from one port to ano<strong>the</strong>r) was <strong>in</strong> such a shared memory.<br />

The problem with <strong>the</strong> shared memory is that <strong>in</strong> <strong>the</strong> worst case it needs to<br />

support accesses from all ports <strong>in</strong> a s<strong>in</strong>gle cell time. Modern switches have<br />

evolved to use cheaper (and slower memory) to build a distributed VC table {<br />

based on <strong>the</strong> assumption that VC label allocations are relatively static (written<br />

only dur<strong>in</strong>g connection setup, read by <strong>the</strong> local port only), and local between<br />

pairs of ports (except po<strong>in</strong>t-to-multipo<strong>in</strong>t VCs which could <strong>in</strong>volve multiple<br />

ports). One disadvantage of <strong>the</strong> distributed memory implementation is that<br />

shar<strong>in</strong>g <strong>in</strong><strong>for</strong>mation between ports via memory is not possible. A solution to this<br />

problem is to haveacheap low speed shared memory (DRAM) <strong>for</strong> stor<strong>in</strong>g shared<br />

tables which are accessed when RM cells are processed. We take advantage of<br />

<strong>the</strong> fact that RM cells on every VC arrive at a frequency of at most one <strong>in</strong><br />

Nrm = 32 cells. Even <strong>in</strong> <strong>the</strong> case when RM cells of multiple VCs arrive<br />

toge<strong>the</strong>r, <strong>the</strong>y need to be processed at a rate much smaller than <strong>the</strong> l<strong>in</strong>k rate.<br />

As mentioned be<strong>for</strong>e, RM cells can be staggered with respect to <strong>the</strong> data stream<br />

as long as <strong>the</strong> sequence <strong>in</strong>tegrity on a VC is ma<strong>in</strong>ta<strong>in</strong>ed.<br />

368

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

Saved successfully!

Ooh no, something went wrong!