[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Beowulf Cluster is very slow. Suggestions needed to increase the speed.



On 2014-10-21 09:42, suresh kannan wrote:
Hello,

Is the code running in the 'single system' the same running in
parallel ?

If i install the software only in the head node (/usr/local/) mpirun
says the software is not available in other nodes and the job
terminates automatically. So, I have installed Gromacs (the software
for simulation) only in head node (/home/mpiuser/)  and it
automatically installed in all other nodes in /home/mpiuser/ (due to
ssh without password, i guess so).

If you are using MPI, you can run the core in parallel within the same machine (1 process for each cpu core). I think this is a god test to compare parallel running time with non-parallel because there isn't network involved.


Is the software using network efficiently ?

I have searched google but i am unable to find out. How can i find out
that the software uses network efficiently. May be can I find that in
openmpi manual?

Debugging...but if you don't write the code this is hard.


Why do you need a layer 3 ? Are cluster's nodes in different
network ?

I learned from few documentations Master node must have two network
cards for this type of cluster but we dont have in our systems. It
seems from the discussions even the USB network card is slow. I dont
know how to add additional network cards to the system. I am running
clusters in the same network (only four systems using a single
router).

You don't need a layer 3 switch (or a router) to connect head node with others nodes. A gigabit switch is enough. Normally the head node have two network cards (A and B) because one (A) connect it with cluster's nodes and the other (B) connect the head node with users' network. So you access the head node by ip address of network card B and the ip address of network card A is used only for internal cluster communication. But this setup is not mandatory, is the one most used.


It seems to me from the discussions that there are three problems

1) we dont have two network cards systems.
No problem.


2) The router i was using (iptime 100Mbps). Although, the speed of
those jobs increased still it is not efficient when i used 1000Mbps
switch compare to a single system.
If possible, use gigabit switch with gigabit network cards.


3) I have used this tutorial
http://flakrat.blogspot.kr/2013/04/how-to-compile-gromacs-461-with-openmpi.html
[2] (for gromacs) to install openmpi, gromacs. Only difference is i
have used gcc compiler instead of intel compiler. That is the only
difference (even they say gcc is make faster code than intel
compilers) . I have read the config (log files for open mpi, gromacs,
fftw libraries etc..) for errors but i found the installations of the
softwares were quite ok. Even if i did some misconfiguration how do i
find out that other than reading the config log files.  Is there are
any other way?
I have never used this software, so I don't know. But if you can compile without errors with gcc it's ok. I personally prefer gcc because is open source.


--

My email was sent by May First/People Link
https://mayfirst.org


Reply to: