Hi,

I am running Fluent 6.1 on Win2K cluster of two nodes each P4 3Gh. I am using rshd for communication. But my case on two nodes runs 5 times slower than single node. I have 100Mbit connection thru a switch. Is there anyone having the same problems or solved already? I'll appreciate too much.

aydin

dear aydin do you have a workgroup or you have defined a domain? do you have win2k adnaced server or just win2k?

I use a cluster of linux os and it works very well. So please tell us more about the architecture of your cluster and the way you have set it up to be able to solve your problem.

Hi david,

I defined a domain on which win2k advanced server is working. the other machine is just win2k.

hi john

I have a domain on which win2k advanced server is working. the other machine is just win2k under this domain.

aydin

So I think you have defined a domain.Don't you think that you could use a workgroup(for example only in win2k).Maybe it is faster. does rshd works perfectly on your system?what are your rshd config's when you 2click on rsh daemon?

hi david,

I was hoping that workgroup would be enough. but rshd deamon must be configured for fluent. and for configuration you have to define a domain. on rshd deamon, before starting the service you have to click on log on tab and click on this account. then you have to choose an account on a domain. (for that reason workgroup is not enough. ) then you start the service. the client node must be under this domain and the same procedure must be followed.

I hope this info is good for you

I can send you a document for that.

Dear Aydin Is your network speed enough good wuthout fluent? (for example when you share a fole or dir) it's probably the problem of your network not fluent.

hi,

I have switch for my cluster. so machines are as if directly connected to each other without noise. my ethernet card is 100 mbit and I can see the lisght on switch saying communication is 100mbit.

I wonder if the problem is with rshd communicator. I dont know what if I use mpi.

why dont you check your rsh? We use it in our office and it doesn't have any problem. so try to test your rshd with a heavy load instruction. my email is again "yoshita4000 at yahoo dot com" I think it's better to make contact with email.

hi misirlioglu,

I want to know the details of your mesh and case used to test your parallel version fluent, such as the size of mesh. I have experience on running fluent on 128-node linux cluster and really found the phenomena you mentioned if the mesh size is small, or the mesh is not properly partitioned and etc. What's more, some cases that reach convergenc in single fluent do diverge.

good luck.

hi,

first I run a 2d case with few elements and saw that it was 5 times slower than serial fluent. then I run a case with 100,000 elements. This time it was not slow that much, although I did not precisely measure the cpu times. So you are right. speed up can be obtained for bigger domains. I used to run the same case on SGI3000 with 8 processors and I almost got linear speed up. I was hoping at least 1.5 times speed up for 2 processors on cluster.

Another thing I used mpi as well instead of rshd. It a bit faster not not remarkable.

aydin

Hello, even i have the same problem and time taken for convergence even further increases as the no. of processor in the cluster increases as compared to single processor. I am also using 100 Mbps switch.

thanks and regards, gaurav