I am using BackupPC with rsync daemons running on the target machines. Some targets are virtual and some are on their own hardware.
A couple of months ago I noticed that one of the VM targets always failed to get a backup, and upon investigation I found that the packet stream contained lots of zero window and keep alive packets.
The failure turned out to be caused by a wrong VM network driver on that particular VM and after resolving that all backups completed. However that caused me to look into it further when I discovered that fixing the driver did not stop the zero window packets from happening and upon further investigation I discovered that ALL backup sessions are seeing lots of zero window packets, then further investigation shows that backups are taking way too long; as long as ten hours for a 35GB backup.
I've googled for zero window errors with rsync and nothing I've found points to anything that could resolve my issue. But how can that be? Am I the only one seeing this issue? Or is it that no one is looking?
I am running CentOS 6 and Fedora 17, all of which have rsync 3.0.9-1, except one, which has 3.0.6-9. I have a Gigibit LAN and iperf shows that there is a solid 900+Mbs throughput in both directions to/from all target machines and the backup server.
Created attachment 8212 [details]
Wireshark packet stream capture (binary tcpdump)
Please use tcp.port==873 as a filter to see an example of results of this bug.