Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
On 10/08/09 17:14, Don Kerr wrote: George, This is an interesting approach although I am guessing the changes would be wide spread and have many performance implications. Am I wrong in this belief? My point here is that if this is going to have as many performance implications as I think it will, it probably makes sense to investigate the potential bigger dual-rail issue and consider the "never share" approach in the larger context. -DON -DON On 10/08/09 11:45, George Bosilca wrote: Don, I think we can do something slightly different that will satisfy everybody. How about a solution where each BTL will define a limit where a message will never be shared with another BTL? We can have two such limits, one for the send protocol and one for the RMA (it will apply either to PUT or GET operations based on the BTL support and PML decision). george. On Oct 8, 2009, at 11:01 , Don Kerr wrote: On 10/07/09 13:52, George Bosilca wrote: Don, The problem is that a particular BTL doesn't have the knowledge about the other selected BTL, so allowing the BTLs to set this limit is not as easy as it sound. However, in the case two identical BTLs are selected and that they are the only ones, this clearly is a better approach. If this parameter is set at the PML level, I can't imagine how we figure out the correct value depending on the BTLs. I see this as a pretty strong restriction. How do we know we set a value that make sense? OK, I now see why setting at btl level is difficult. And for the case of multiple btls which are also different component types, however unlikely that is, a pml setting will not be optimal for both. -DON george. On Oct 7, 2009, at 10:19 , Don Kerr wrote: George, Were you suggesting that the proposed new parameter "max_rdma_single_rget" be set by the individual btls similar to "btl_eager_limit"? Seems to me to that is the better approach if I am to move forward with this. -DON On 10/06/09 11:14, Don Kerr wrote: I agree there is probably a larger issue here and yes this is somewhat specific but where as OB1 appears to have multiple protocols depending on the capabilities of the BTLs I would not characterize as an IB centric problem. Maybe OB1 RDMA problem. There is a clear benefit from modifying this specific case. Do you think its not worth making incremental improvements while also attacking a potential bigger issue? -DON On 10/06/09 10:52, George Bosilca wrote: Don, This seems a very IB centric problem (and solution) going up in the PML. Moreover, I noticed that independent on the BTL we have some problems with the multi-rail performance. As an example on a cluster with 3 GB cards we get the same performance is I enable 2 or 3. Didn't had time to look into the details, but this might be a more general problem. george. On Oct 6, 2009, at 09:51 , Don Kerr wrote: I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
George, This is an interesting approach although I am guessing the changes would be wide spread and have many performance implications. Am I wrong in this belief? -DON On 10/08/09 11:45, George Bosilca wrote: Don, I think we can do something slightly different that will satisfy everybody. How about a solution where each BTL will define a limit where a message will never be shared with another BTL? We can have two such limits, one for the send protocol and one for the RMA (it will apply either to PUT or GET operations based on the BTL support and PML decision). george. On Oct 8, 2009, at 11:01 , Don Kerr wrote: On 10/07/09 13:52, George Bosilca wrote: Don, The problem is that a particular BTL doesn't have the knowledge about the other selected BTL, so allowing the BTLs to set this limit is not as easy as it sound. However, in the case two identical BTLs are selected and that they are the only ones, this clearly is a better approach. If this parameter is set at the PML level, I can't imagine how we figure out the correct value depending on the BTLs. I see this as a pretty strong restriction. How do we know we set a value that make sense? OK, I now see why setting at btl level is difficult. And for the case of multiple btls which are also different component types, however unlikely that is, a pml setting will not be optimal for both. -DON george. On Oct 7, 2009, at 10:19 , Don Kerr wrote: George, Were you suggesting that the proposed new parameter "max_rdma_single_rget" be set by the individual btls similar to "btl_eager_limit"? Seems to me to that is the better approach if I am to move forward with this. -DON On 10/06/09 11:14, Don Kerr wrote: I agree there is probably a larger issue here and yes this is somewhat specific but where as OB1 appears to have multiple protocols depending on the capabilities of the BTLs I would not characterize as an IB centric problem. Maybe OB1 RDMA problem. There is a clear benefit from modifying this specific case. Do you think its not worth making incremental improvements while also attacking a potential bigger issue? -DON On 10/06/09 10:52, George Bosilca wrote: Don, This seems a very IB centric problem (and solution) going up in the PML. Moreover, I noticed that independent on the BTL we have some problems with the multi-rail performance. As an example on a cluster with 3 GB cards we get the same performance is I enable 2 or 3. Didn't had time to look into the details, but this might be a more general problem. george. On Oct 6, 2009, at 09:51 , Don Kerr wrote: I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
Don, I think we can do something slightly different that will satisfy everybody. How about a solution where each BTL will define a limit where a message will never be shared with another BTL? We can have two such limits, one for the send protocol and one for the RMA (it will apply either to PUT or GET operations based on the BTL support and PML decision). george. On Oct 8, 2009, at 11:01 , Don Kerr wrote: On 10/07/09 13:52, George Bosilca wrote: Don, The problem is that a particular BTL doesn't have the knowledge about the other selected BTL, so allowing the BTLs to set this limit is not as easy as it sound. However, in the case two identical BTLs are selected and that they are the only ones, this clearly is a better approach. If this parameter is set at the PML level, I can't imagine how we figure out the correct value depending on the BTLs. I see this as a pretty strong restriction. How do we know we set a value that make sense? OK, I now see why setting at btl level is difficult. And for the case of multiple btls which are also different component types, however unlikely that is, a pml setting will not be optimal for both. -DON george. On Oct 7, 2009, at 10:19 , Don Kerr wrote: George, Were you suggesting that the proposed new parameter "max_rdma_single_rget" be set by the individual btls similar to "btl_eager_limit"? Seems to me to that is the better approach if I am to move forward with this. -DON On 10/06/09 11:14, Don Kerr wrote: I agree there is probably a larger issue here and yes this is somewhat specific but where as OB1 appears to have multiple protocols depending on the capabilities of the BTLs I would not characterize as an IB centric problem. Maybe OB1 RDMA problem. There is a clear benefit from modifying this specific case. Do you think its not worth making incremental improvements while also attacking a potential bigger issue? -DON On 10/06/09 10:52, George Bosilca wrote: Don, This seems a very IB centric problem (and solution) going up in the PML. Moreover, I noticed that independent on the BTL we have some problems with the multi-rail performance. As an example on a cluster with 3 GB cards we get the same performance is I enable 2 or 3. Didn't had time to look into the details, but this might be a more general problem. george. On Oct 6, 2009, at 09:51 , Don Kerr wrote: I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
On 10/07/09 13:52, George Bosilca wrote: Don, The problem is that a particular BTL doesn't have the knowledge about the other selected BTL, so allowing the BTLs to set this limit is not as easy as it sound. However, in the case two identical BTLs are selected and that they are the only ones, this clearly is a better approach. If this parameter is set at the PML level, I can't imagine how we figure out the correct value depending on the BTLs. I see this as a pretty strong restriction. How do we know we set a value that make sense? OK, I now see why setting at btl level is difficult. And for the case of multiple btls which are also different component types, however unlikely that is, a pml setting will not be optimal for both. -DON george. On Oct 7, 2009, at 10:19 , Don Kerr wrote: George, Were you suggesting that the proposed new parameter "max_rdma_single_rget" be set by the individual btls similar to "btl_eager_limit"? Seems to me to that is the better approach if I am to move forward with this. -DON On 10/06/09 11:14, Don Kerr wrote: I agree there is probably a larger issue here and yes this is somewhat specific but where as OB1 appears to have multiple protocols depending on the capabilities of the BTLs I would not characterize as an IB centric problem. Maybe OB1 RDMA problem. There is a clear benefit from modifying this specific case. Do you think its not worth making incremental improvements while also attacking a potential bigger issue? -DON On 10/06/09 10:52, George Bosilca wrote: Don, This seems a very IB centric problem (and solution) going up in the PML. Moreover, I noticed that independent on the BTL we have some problems with the multi-rail performance. As an example on a cluster with 3 GB cards we get the same performance is I enable 2 or 3. Didn't had time to look into the details, but this might be a more general problem. george. On Oct 6, 2009, at 09:51 , Don Kerr wrote: I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
Don, The problem is that a particular BTL doesn't have the knowledge about the other selected BTL, so allowing the BTLs to set this limit is not as easy as it sound. However, in the case two identical BTLs are selected and that they are the only ones, this clearly is a better approach. If this parameter is set at the PML level, I can't imagine how we figure out the correct value depending on the BTLs. I see this as a pretty strong restriction. How do we know we set a value that make sense? george. On Oct 7, 2009, at 10:19 , Don Kerr wrote: George, Were you suggesting that the proposed new parameter "max_rdma_single_rget" be set by the individual btls similar to "btl_eager_limit"? Seems to me to that is the better approach if I am to move forward with this. -DON On 10/06/09 11:14, Don Kerr wrote: I agree there is probably a larger issue here and yes this is somewhat specific but where as OB1 appears to have multiple protocols depending on the capabilities of the BTLs I would not characterize as an IB centric problem. Maybe OB1 RDMA problem. There is a clear benefit from modifying this specific case. Do you think its not worth making incremental improvements while also attacking a potential bigger issue? -DON On 10/06/09 10:52, George Bosilca wrote: Don, This seems a very IB centric problem (and solution) going up in the PML. Moreover, I noticed that independent on the BTL we have some problems with the multi-rail performance. As an example on a cluster with 3 GB cards we get the same performance is I enable 2 or 3. Didn't had time to look into the details, but this might be a more general problem. george. On Oct 6, 2009, at 09:51 , Don Kerr wrote: I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
George, Were you suggesting that the proposed new parameter "max_rdma_single_rget" be set by the individual btls similar to "btl_eager_limit"? Seems to me to that is the better approach if I am to move forward with this. -DON On 10/06/09 11:14, Don Kerr wrote: I agree there is probably a larger issue here and yes this is somewhat specific but where as OB1 appears to have multiple protocols depending on the capabilities of the BTLs I would not characterize as an IB centric problem. Maybe OB1 RDMA problem. There is a clear benefit from modifying this specific case. Do you think its not worth making incremental improvements while also attacking a potential bigger issue? -DON On 10/06/09 10:52, George Bosilca wrote: Don, This seems a very IB centric problem (and solution) going up in the PML. Moreover, I noticed that independent on the BTL we have some problems with the multi-rail performance. As an example on a cluster with 3 GB cards we get the same performance is I enable 2 or 3. Didn't had time to look into the details, but this might be a more general problem. george. On Oct 6, 2009, at 09:51 , Don Kerr wrote: I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel ___ devel mailing list de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
[OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)
I intend to make the change suggested in this ticket to the trunk. The change does not impact single rail, tested with openib btl, case and does improve dual rail case. Since it does involve performance and I am adding a OB1 mca parameter just wanted to check if anyone was interested or had an issue with it before I committed the change. -DON
