On 3/22/2012 3:43 PM, jukka.tuomi...@finndesign.fi wrote: > > Hi all, > > here are some ping results outside LAN: > > 1.62 1.80 1.62 2.33 1.69 1.64 1.75 2.35 5.77 3.85 4.69 2.73 ... > > The internet speed is the same 100/5. > > I just moved ~20M/~2000 files > AFS to SSH server ~220KB/s > SSH server to AFS ~80KB/s > > The client was in WAN this time, whereas both servers where in the same > LAN (not sure it they knew it :). > > br, jukka
On average your files are 10K. 10K is smaller than both the RPC chunksize and the window size. For each read or write of a file there will be a single RPC. The big difference is going to occur due to directory operations. When writing to AFS you are going to issue a CreateFile RPC for each file entry. When you read from AFS you are going to issue one InlineBulkStatus RPC for each 50 entries in the directory. In other words, 2000 RPCs when writing and 400 RPCs when reading. Ping times are not representative of baseline RPC latency nor will they include client side file system overhead for the two different execution paths. You will see very different results if you read and write a single 20MB file or even a 1GB file. For maximum throughput with large files and the OpenAFS master source code branch you can use a window size of 254 and a chunksize of 8MB. Increasing the window size and the chunksize have no impact when the file sizes are small as they are in your test. Jeffrey Altman
signature.asc
Description: OpenPGP digital signature