Wednesday, November 29, 2006

Week of November 23-29

I did the literature survey of my research area including GridFTPXIO, GridFTP, gLite, RLS.

Power point presentation about my research are can downloaded from the following link:

Tuesday, November 14, 2006

Week of November 9-15

Additional tests have been performed on LAN and WAN. Best achieved network transfer speeds are around 71 Mbps between Bloomington and Indianapolis and 19 Mbps between Bloomington and San Diego with four source and one client. To increase the performance gain the followings parameters have been modified.

1-) Different piece size (from 256 KB-4 MB)
2-) Different block size (from 16KB-512KB)

No performance gain has been observed. In my opinion, there are already overhead and synchronization issues in GridTorrent and adding new features to increase transfer speed has adverse effect on its data transfer performace.

Wednesday, November 08, 2006

LAN and WAN tests (II)

LAN test result:

Sources were deployed on gridfarm machines and client was deployed on complexity machine. The average transfer speed is around 70 Mbps. For the same situation -4 parallel streams- PTCP transfer speed is around 80 Mbps. Total file size is around 185MB and the bandwidth usage of each source:

source1source2source3source4
44MB53MB47MB41MB


WAN test result:

Like LAN tests, sources were deployed on gridfarm machines and client was deployed on pipeline3 machine, San Diego. The same performance pattern in LAN has been observed, too. The average transfer speed for Gridtorrent around is 17 Mbps and for PTCP -4 parallel streams- is around 27 Mbps.

source1source2source3source4
52MB45MB42MB46MB


Observations:

In both LAN and WAN there is an about 10 Mbps drop observed on Gridtorrent transfer. I think this drop is normal due to nature of Gridtorrent. In PTCP, when the communication is set, there is not communication overhead any more. Just data is transfered. However, in Gridtorrent, there is a request overhead for each piece and blocks of each pieces, and integrity of each piece is verified by hash-code checking. This communication and hash-code checking overhead degrades overall data transfer. Even though, the drop in data transfer speed, sources bandwidth usage seems balanced which is the main point of Bittorrent inspired Gridtorrent.

Many Gridtorrent parameters were changed to see how its performance behaves under those changes. But there is no significant effect on performance gain of Gridtorrent.

11/08/2006, Bloomington