Archives


- Beowulf
- Beowulf Announce
- Scyld-users
- Beowulf on Debian

mini-cluster network delays

Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.

Search

Kris Thielemans kris.thielemans at ic.ac.uk
Wed Jan 17 07:51:22 PST 2001


Hi,

we have really funny timings when getting data from one Linux machine to
another. I did the timings with ftp (because that reports them), but I have
the impression the same problems exist when doing regular cp etc. Also,
because I use ftp, I think I minimise NIS access (I don't need to check
mount points etc.)
Hi,

I have some delays in network access between different machines in our
mini-cluster. This is not really a beowulf (yet), but I hope somebody can
help me how to find out what's going wrong.

Our cluster has 4 identical Linux PCs (SuSE 6.3), each of them with  100Mbit
Ethernet connection to an Ethernet switch (and from the switch also a 10Mbit
to the 'outside').


We have really funny timings when getting data from one Linux machine to
another. I did the timings with ftp (because that reports them), but I have
the impression the same problems exist when doing regular cp etc. Also,
because I use ftp, I think I minimise NIS access (I don't need to check
mount points etc.)

So I'm ftp-ing a file of 5.7 MB (binary mode). Below is are some examples of
timings of ftp-ing this file between the different machines (called pp1-pp4)

from	to	time (secs)
pp2	pp1	6.5
pp1	pp2	20
pp1	pp3	1.5
pp3	pp1	1.5
pp3	pp2	21
pp2	pp3	6.5
pp1	pp4	21
pp4	pp1	6.5
pp2	pp4	7
pp4	pp2	6.5
pp3	pp4	21.5

as comparison:
pp3	pp3		0.14 (from 1 machine to itself)
pp1	outside	17.5 (depends on network traffic heavily)
outside pp1		7.5

Obviously we want to have the performance of pp1<->pp3 everywhere.

Any ideas where this can come from or how to find out what causes this?

Thanks
Kris

PS:
(One fairly nasty thing is that users and groups are defined on a NIS server
which is 'outside'. We want to rememdy this in the future, but had some
problems having a local NIS server on one machine, which would still know
about the outside NIS server. Anyway. I hope this i snot important right
now.)

Kris Thielemans
MRC Cyclotron Unit,
Hammersmith Hospital,
DuCane Rd,London W12 0NN, United Kingdom

NEW web site address:
http://www.cu.mrc.ac.uk/~kris






More information about the Beowulf mailing list