In this paper, we present five new sets of performance results.
1) We examine the performance of affinity-based scheduling when
a large number of streams are concurrently supported. In our ear.-
lier work, we fixed the number of concurrent streams at eight. The
benefit of affinity-based scheduling is maintained for both PLP
and CLP. 2) We use affinity-based scheduling in send-side proto.-
col processing, motivated by the communication requirements of
large-scale server applications such as video-on-demand. We find
the scheduling technique performs well for send-side processing,
in agreement with our earlier receive-side results. 3) We consider
the impact of data-touching protocol-processing operations, for
both send-side and receive-side processing. Most network implementations
copy packet data, an operation which perturbs cached
protocol state and can dominate packet processing time. We experimentally
measure the impact of copying uncached packet data
on packet processing time, and incorporate the overhead into our
results using an empirical UDPIIPRDDI packet-size distribution.
The benefit of affinity-based schedulingremains significant. 4) We
evaluate performance as a function of stream burstiness and sourc1E
locality, two well-known properties of network traffic [9, 141, by
modeling individual streams with the Packet-Train source model
[9]. We find that the performance of affinity-based scheduling
is relatively insensitive to these source characteristics. 5) Wit
explore a technique for improving the caching behavior and available
packet-level concurrency under CLP. By matching the number
of “independent protocol stacks” to the number of admitted
streams, the performance of CLP improves dramatically, due tto