Master Dissertation

SCTP performance test - protocol de aplicação SMB/CIFS

Scenario: COP as server, SOLDIER as client. Two network sceneries were used:

100Mbps Ethernet;
1Mbps with traffic shaping at reception, without fixed packet losses, 30ms latency, 0.1% packet duplication.

Application adapted for the tests: Samba version 3.04. Client program was smbclient that is part of Samba package.

Throughput test was simply the transmission of a big file (100 million bytes at 100Mbps, 1 million bytes at 1Mbps).

Latency test consisted of repeated execution of two commands: transmission of a 100-byte sized file, followd by immediate deletion. At 100Mbps, this command pair was repeated 5000 times; at 1Mbps, it was repeated only 100 times.

This test does not express throughput and latency in the usual units. We just measured the "wall clock" elapsed during test execution. This avoided further instrumentation of applications, and gave an "end user" perspective to the test, since end user just perceives "wall clock" while it wait (im)patiently until some operation is completed.

Network scenario	TCP - throughput	SCTP - throughput	TCP - latency	SCTP - latency
100Mbps	11,54	12,39	6,12	7,13
1Mbps	8,53	11,51	30,02	30,09

Table values in "wall clock" seconds (see scenario remarks, in the top of this page).

It was expected that simple substitution from TCP to SCTP would increase SMB performance, since it uses messsages smaller than 64kpbs. Tests showed that it is not sufficient. Possibly a deeper adaptation of SMB to SCTP (e.g. disabling message separation routines since SCTP supplies that) could improve SCTP results.