Packet-Level Parallelism

Packet-Level Parallelism

Times given are in megabits per second for several protocol stacks. Our baseline
Internet stack consists of TCP/IP/FDDI, representing
protocol processing without any security. A second stack is
an Internet stack with MD5 between TCP and IP, representing
the work done for an application that requires authentication
and integrity but no confidentiality. Our third stack
uses DES above TCP and MD5 below TCP, which supports
both confidentiality and integrity. Our fourth stack is the
same as the third, except that we use triple-DES instead of
DES.

These throughputs were measured on our 12-processor
Challenge machine, using a single TCP connection with 4
KB packets. For these and all subsequent graphs, each data
point is the average of 10 runs, where a run consists of
measuring the steady-state throughput for 30 seconds, after
an initial 30 second warmup period. Throughput graphs
include 90 percent confidence intervals.


The baseline speed for the send-side TCP stack is roughly 138 Mbits/sec. Adding MD5 to the stack reduces throughput by nearly an order of magnitude, to
a mere 18 Mbits/sec2 . Adding DES on top of TCP reduces
throughput nearly 2 orders of magnitude, to 4.6 Mbits/sec.
Using Triple-DES is 3 times slower at 1.5 Mbits/sec.
Figure 4 shows the corresponding relative speedup for
the send-side tests, where speedup is throughput normalized
relative to the uniprocessor throughput for the appropriate
stack. The theoretical ideal linear speedup is included for
comparison. Previous work [3, 24] has shown limited performance
gains when using packet-level parallelism for a
single TCP connection, barring any other protocol processing,
and this is reflected by the baseline TCP/IP stack’s
minimal speedup. This is because manipulating a TCP connection’s
state is large relative to the IP and FDDI processing
and must occur inside a single locked, serial component. Of
course, throughput can be improved by using multiple connections.

However, as more compute-intensive cryptographic protocols
are used, while the throughput goes down, the relative
speedup improves. For example, the MD5 stack achieves a
speedup of 8 with 12 processors, and the DES and Triple-
DES stacks produce very close to linear speedup