Bandwidth Test Notes

Introduction

In a distributed processing environment transferring data over the wire is sometimes unavoidable. Since network bandwidth is still relatively expensive and since computing power and disk space are cheap commodities, the accepted wisdom is so limit data transfers to the absolute minimum.

This page gathers test results and information pointing out some strategies for those situations when data transfers are unavoidable.

Data Transfer Speed Tests

Preliminary results from data transfer tests between vonc2 and vodev1 are summarized below. Four files of approximately 300 MB were transferred from vodev1 to vonc2 using a number of different tools. The transfer speed reported below for each tool are the average of the four transfer speeds.

Tool	Min Speed (MB/s)	Max Speed (MB/s)	Avg. Speed (% Max B/W)	Avg. Speed (MB/s)
iperf 1.7.0	24.2500	24.3750	-	24.2813
rsync	14.2457	14.9646	59	14.4686
scp	13.7893	14.0577	57	13.9259
bbftp	14.0634	16.1186	62	15.1549

The bandwidth between the two machines (vodev1 and vonc2) was measured with iperf to be approximately 24 MB/s.

bbFTP was invoked (on vodev1) with the command:

    % bbftp -u vosw -e "setnbstream 10; put <str file> <dst file>" vonc2 -E "bbftpd -s" >& /dev/null

The number of streams (10) was chosen at random but it appears to be the maximum the tool supports.

Clearly, bbFTP appears to be faster than both scp and rsync. Its configuration is not trivial and the choice of parameters will require some testing and network characterization. Possibly, the exact parameter set for bbFTP will depend on the pair of machines the transfer takes place between. Moreover the scatter in speed figures for bbFTP is quite bothersome.

A note of caution is in order: no particular care had been taken in assuring that the two machines's network connections were free (only that their CPU was free, which has some relevance to the tests above, but not too nuch relevance).

-- FrancescoPierfederici - 01 Feb 2006