Specifications

3.3 Memory and CPU Testing
The first tests which incorporated socket-based communication aimed at verifying that the CPU
and memory could maintain sufficient transfer rates. By creating a socket connection to “localhost”,
the test was able to send UDP datagrams from one part of the data path and have them received
in another. Because the Linux so cket library copies the data from application memory into system
memory when sending, and from system memory to application memory when receiving, this test
effectively measured the maximum transfer rate of the CPU and memory under the same usage
pattern as the MCS-DR requires. Like the hard drive benchmarking tests, packets consisting of a
serial identifier, and a series of 8-bit counter values were used. In this case, however, the exhaus-
tive checking of the entire packet would have perturbed the results of the test, and only the serial
identifier was verified. The noted maximum transfer rate was 465 MiB/s with 500 byte packets.
The performance limitation in this scheme comes not from the overall data rate, but rather it exists
as a relationship between the size of the packets and the number of packets transfers required per
second. With arbitrarily large packets, the limit approaches the maximum memory bandwidth of
the system, and as the packets get smaller, the overhead of kernel IO MMU
4
calls required to copy
the data dominates. Since the minimum packet size of TBN, TBW, and DRX packets is the TBN
packet size of 1008 bytes, 500 bytes was arbitrarily chosen as being sufficiently smaller than the
packets of interest as to ensure that success of the test would imply that any larger sized packet
transfer would also meet the data rate requirements.
3.4 Network Performance
Testing has been performed which verified the network adapter’s ability to meet the system re-
quirements. Two sets of tests were performed. The first, preliminary tests of hardware driver and
functionality were included as part of the driver package from by Myricom. The included loopback
test measures transfer speeds of the network adapter. This test was run for four hours and for ten
hours, with the transfer rate approaching 4 Gib/s in both cases. The second set of tests were the
duration tests described in the next section. The successes of both series of tests are sufficient to
validate the network hardware selection because they utilize all of the hardware components essential
to the MCS-DR PC’s core functionality.
4
Input Output Memory Management Unit
10