Re: [gmx-users] Gromacs performance on virtual servers
On Fri, Jul 25, 2014 at 1:51 AM, Szilárd Páll pall.szil...@gmail.com wrote: Hi In general, virtualization will always have an overhead, but if done well, the performance should be close to that of bare metal. However, for GROMACS the ideal scenario is exclusive host access (including hypervisor) and thread affinities which will both depend on the hypervisor configuration. Hence, if you can, you should try to get access to virtual hosts that fully utilize a compute node and do not share it with others. Definitely. On Fri, Jul 25, 2014 at 12:31 AM, Mark Abraham mark.j.abra...@gmail.com wrote: Hi, Except for huge simulation systems, GROMACS performance past a single node is dominated by network latency, so unless you can extract a promise that any multi-node runs will have Infiniband-quality latency (because the nodes are physically in the same room, and on Infiniband) you can forget about doing multi-node MD on such a system. Two remarks: * With a slow network the only parallelization you can potentially make use of is multi-sim, unless your environment is so could-y that some nodes can have tens to hundreds of ms latency which can kill even you multi-sim performance (depending on how fast each simulation is and how often do they sync). I would not encourage multi-sim on such a setup, unless you actually want replica exchange. The multi-sim implementation unnecessarily syncs simulations every min(nstlist,nstcalcenergy,nstreplex) step, so that might be ~tens of times per second. Unnecessary multi-sim is good for pretending you are doing a big parallel calculation to get access to a large chunk of a machine, but this is not really the case here. * I've seen several claims that *good* 10/40G Ethernet can get close to IB even in latency, even for MD, and even for GROMACS, e.g: http://goo.gl/JrNxKf, http://goo.gl/t0z15f Interesting, thanks. Mark Cheers, -- Szilárd Mark On Thu, Jul 24, 2014 at 10:54 PM, Elton Carvalho elto...@if.usp.br wrote: Dear Gromacs Users, My former university is focusing on cloud computing instead of physical servers, so research groups are now expected to buy virtual servers from the university coloud instead of buying their own clusters. The current setup employs Xeon E7- 2870 servers and there is an university-wide virtual cluster with 50 virtual servers each with 10 CPUs. Does anyone here have information on gromacs performance on this kind of infrastructure? Should I expect big issues? One thing that comes to mind is that the CPUs may not necessarily be in the same physical server, rack, or even datacenter (their plan is to decentralize the colocation), so network latency may be higher than the traditional setup, which may affect scaling. Does this argument make sense or am I missing something on cloud management 101? Cheers. -- Elton Carvalho Departamento de Física Universidade Federal do Paraná -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org.
Re: [gmx-users] Gromacs performance on virtual servers
On Fri, Jul 25, 2014 at 12:33 PM, Mark Abraham mark.j.abra...@gmail.com wrote: On Fri, Jul 25, 2014 at 1:51 AM, Szilárd Páll pall.szil...@gmail.com wrote: Hi In general, virtualization will always have an overhead, but if done well, the performance should be close to that of bare metal. However, for GROMACS the ideal scenario is exclusive host access (including hypervisor) and thread affinities which will both depend on the hypervisor configuration. Hence, if you can, you should try to get access to virtual hosts that fully utilize a compute node and do not share it with others. Definitely. On Fri, Jul 25, 2014 at 12:31 AM, Mark Abraham mark.j.abra...@gmail.com wrote: Hi, Except for huge simulation systems, GROMACS performance past a single node is dominated by network latency, so unless you can extract a promise that any multi-node runs will have Infiniband-quality latency (because the nodes are physically in the same room, and on Infiniband) you can forget about doing multi-node MD on such a system. Two remarks: * With a slow network the only parallelization you can potentially make use of is multi-sim, unless your environment is so could-y that some nodes can have tens to hundreds of ms latency which can kill even you multi-sim performance (depending on how fast each simulation is and how often do they sync). I would not encourage multi-sim on such a setup, unless you actually want replica exchange. The multi-sim implementation unnecessarily syncs simulations every min(nstlist,nstcalcenergy,nstreplex) step, so that might be ~tens of times per second. Unnecessary multi-sim is good for pretending you are doing a big parallel calculation to get access to a large chunk of a machine, but this is not really the case here. Good point, I fully agree! What I meant was that one can use quite efficiently even cheap Ethernet networks for multi-sim type workloads where there is genuine need for communication between simulations (but not too often). For instance, replica exchange type runs work quite OK. -- Szilárd * I've seen several claims that *good* 10/40G Ethernet can get close to IB even in latency, even for MD, and even for GROMACS, e.g: http://goo.gl/JrNxKf, http://goo.gl/t0z15f Interesting, thanks. Mark Cheers, -- Szilárd Mark On Thu, Jul 24, 2014 at 10:54 PM, Elton Carvalho elto...@if.usp.br wrote: Dear Gromacs Users, My former university is focusing on cloud computing instead of physical servers, so research groups are now expected to buy virtual servers from the university coloud instead of buying their own clusters. The current setup employs Xeon E7- 2870 servers and there is an university-wide virtual cluster with 50 virtual servers each with 10 CPUs. Does anyone here have information on gromacs performance on this kind of infrastructure? Should I expect big issues? One thing that comes to mind is that the CPUs may not necessarily be in the same physical server, rack, or even datacenter (their plan is to decentralize the colocation), so network latency may be higher than the traditional setup, which may affect scaling. Does this argument make sense or am I missing something on cloud management 101? Cheers. -- Elton Carvalho Departamento de Física Universidade Federal do Paraná -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! *
Re: [gmx-users] Gromacs performance on virtual servers
Hi, Except for huge simulation systems, GROMACS performance past a single node is dominated by network latency, so unless you can extract a promise that any multi-node runs will have Infiniband-quality latency (because the nodes are physically in the same room, and on Infiniband) you can forget about doing multi-node MD on such a system. Mark On Thu, Jul 24, 2014 at 10:54 PM, Elton Carvalho elto...@if.usp.br wrote: Dear Gromacs Users, My former university is focusing on cloud computing instead of physical servers, so research groups are now expected to buy virtual servers from the university coloud instead of buying their own clusters. The current setup employs Xeon E7- 2870 servers and there is an university-wide virtual cluster with 50 virtual servers each with 10 CPUs. Does anyone here have information on gromacs performance on this kind of infrastructure? Should I expect big issues? One thing that comes to mind is that the CPUs may not necessarily be in the same physical server, rack, or even datacenter (their plan is to decentralize the colocation), so network latency may be higher than the traditional setup, which may affect scaling. Does this argument make sense or am I missing something on cloud management 101? Cheers. -- Elton Carvalho Departamento de Física Universidade Federal do Paraná -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org.
Re: [gmx-users] Gromacs performance on virtual servers
On Fri, Jul 25, 2014 at 1:51 AM, Szilárd Páll pall.szil...@gmail.com wrote: Hi In general, virtualization will always have an overhead, but if done well, the performance should be close to that of bare metal. However, for GROMACS the ideal scenario is exclusive host access (including hypervisor) and thread affinities which will both depend on the hypervisor configuration. Hence, if you can, you should try to get access to virtual hosts that fully utilize a compute node and do not share it with others. On Fri, Jul 25, 2014 at 12:31 AM, Mark Abraham mark.j.abra...@gmail.com wrote: Hi, Except for huge simulation systems, GROMACS performance past a single node is dominated by network latency, so unless you can extract a promise that any multi-node runs will have Infiniband-quality latency (because the nodes are physically in the same room, and on Infiniband) you can forget about doing multi-node MD on such a system. Two remarks: * With a slow network the only parallelization you can potentially *inter-node parallelizaion make use of is multi-sim, unless your environment is so could-y that some nodes can have tens to hundreds of ms latency which can kill even you multi-sim performance (depending on how fast each simulation is and how often do they sync). * I've seen several claims that *good* 10/40G Ethernet can get close to IB even in latency, even for MD, and even for GROMACS, e.g: http://goo.gl/JrNxKf, http://goo.gl/t0z15f Cheers, -- Szilárd Mark On Thu, Jul 24, 2014 at 10:54 PM, Elton Carvalho elto...@if.usp.br wrote: Dear Gromacs Users, My former university is focusing on cloud computing instead of physical servers, so research groups are now expected to buy virtual servers from the university coloud instead of buying their own clusters. The current setup employs Xeon E7- 2870 servers and there is an university-wide virtual cluster with 50 virtual servers each with 10 CPUs. Does anyone here have information on gromacs performance on this kind of infrastructure? Should I expect big issues? One thing that comes to mind is that the CPUs may not necessarily be in the same physical server, rack, or even datacenter (their plan is to decentralize the colocation), so network latency may be higher than the traditional setup, which may affect scaling. Does this argument make sense or am I missing something on cloud management 101? Cheers. -- Elton Carvalho Departamento de Física Universidade Federal do Paraná -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org.
Re: [gmx-users] Gromacs performance on virtual servers
Hi In general, virtualization will always have an overhead, but if done well, the performance should be close to that of bare metal. However, for GROMACS the ideal scenario is exclusive host access (including hypervisor) and thread affinities which will both depend on the hypervisor configuration. Hence, if you can, you should try to get access to virtual hosts that fully utilize a compute node and do not share it with others. On Fri, Jul 25, 2014 at 12:31 AM, Mark Abraham mark.j.abra...@gmail.com wrote: Hi, Except for huge simulation systems, GROMACS performance past a single node is dominated by network latency, so unless you can extract a promise that any multi-node runs will have Infiniband-quality latency (because the nodes are physically in the same room, and on Infiniband) you can forget about doing multi-node MD on such a system. Two remarks: * With a slow network the only parallelization you can potentially make use of is multi-sim, unless your environment is so could-y that some nodes can have tens to hundreds of ms latency which can kill even you multi-sim performance (depending on how fast each simulation is and how often do they sync). * I've seen several claims that *good* 10/40G Ethernet can get close to IB even in latency, even for MD, and even for GROMACS, e.g: http://goo.gl/JrNxKf, http://goo.gl/t0z15f Cheers, -- Szilárd Mark On Thu, Jul 24, 2014 at 10:54 PM, Elton Carvalho elto...@if.usp.br wrote: Dear Gromacs Users, My former university is focusing on cloud computing instead of physical servers, so research groups are now expected to buy virtual servers from the university coloud instead of buying their own clusters. The current setup employs Xeon E7- 2870 servers and there is an university-wide virtual cluster with 50 virtual servers each with 10 CPUs. Does anyone here have information on gromacs performance on this kind of infrastructure? Should I expect big issues? One thing that comes to mind is that the CPUs may not necessarily be in the same physical server, rack, or even datacenter (their plan is to decentralize the colocation), so network latency may be higher than the traditional setup, which may affect scaling. Does this argument make sense or am I missing something on cloud management 101? Cheers. -- Elton Carvalho Departamento de Física Universidade Federal do Paraná -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org. -- Gromacs Users mailing list * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting! * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists * For (un)subscribe requests visit https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-requ...@gromacs.org.