home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!cs.utexas.edu!uwm.edu!caen!batcomputer!ghost.dsi.unimi.it!univ-lyon1.fr!chx400!sicsun!sic!possoz
- From: possoz@sic.epfl.ch (Anne Possoz EPFL-SIC/SL)
- Newsgroups: comp.protocols.tcp-ip
- Subject: RPC vs sockets performance and network influence
- Keywords: RPC socket performance
- Message-ID: <1992Nov19.114117@sic.epfl.ch>
- Date: 19 Nov 92 10:41:16 GMT
- Sender: news@sicsun.epfl.ch
- Reply-To: possoz@sic.epfl.ch
- Organization: Ecole Polytechnique Federale de Lausanne
- Lines: 34
-
- Working a lot on transfer of huge arrays across different computers,
- I reached the following conclusions.
-
- 1) using Sun RPC (with underlying XDR), I can handle .2 to .6 Mbytes/sec
- depending on the platform on which I am working (here the CPU time
- is converted in Mbytes/sec); this has to be compared to the network
- which can handle around 1. Mbytes/sec for ethernet.
- So the CPU is slower than the network. The reason for this is that
- the XDR package is not at all optimized for this kind of problems.
-
- 2) using sockets, I can handle 3. to 10. Mbytes/sec instead. So now the
- network is my limitation. Moreover, I noticed that if data are going
- from FFDI network to ethernet network, I reach that performance only
- if I decrease the SO_RCVBUF size to 8 kbytes (the gain is a factor of 3
- to 4 if SO_RCVBUF is set to 8 kbytes compared to 32 kbytes).
- Note that by default, SO_RCVBUF is 4K on SUN, 8K on HP, 32K on crays and
- 61K on SG. There is also some interest to increase the SO_SNDBUF to a value
- of 32K.
-
- Thus I have some questions to who migth help :
-
- - where can I find any information on performances?
-
- - does this means that any application using RPC are not suitable for large
- transfer of data? And, as far as I can read, NFS is based on RPC; does that
- mean that NFS is not optimized?
-
- - do you know of any other newsgroup where I could read on that subject?
-
- Anne
- --
-
- Anne Possoz
- possoz@sic.epfl.ch
-