Virtualisation has obvious benefits for much commercial IT, where existing servers often have utilisation rates of 10% or less. It's less clear whether virtualisation is so useful for High-Performance Computing (HPC), where systems are often kept running at utilisation rates above 90%. Nevertheless, there are potential benefits from adopting some aspects of virtualisation in these environments. The question is, do the benefits outweigh the cost in performance?
This was the topic of yesterday's workshop on System Level Virtualisation for HPC (which was part of EuroSys 2008). The workshop was rather small but I did learn quite a bit.
A couple of talks investigated the performance hits of using virtualised OS's instead of addressing the hardware directly. This varied, depending on the type of application; if IO was minimal, the slowdown was minimal too. An "average" slowdown seemed to be on the order of 8%.
Samuel Thibault of Xensource looked at ways of using the flexibility of a hypervisor to implement just those parts of an operating system that are absolutely necessary - ignoring more general-purpose facilities such as virtual memory, pre-emptive threads and the like. Essentially, this was using Xen as a "micro-kernel done right".
Larry Rudolph, who works for VMWare but was here speaking with his MIT hat, gave a scenario where virtualisation might benefit the HPC user. Some problems require massive machines that researchers only get access to maybe twice a year. The example Larry gave was modelling global ocean currents. If the researchers can save the messages sent between regions of this model, they can later run more detailed analyses on one particular region on a much smaller machine, using the recorded messages to playback the effect of the rest of the simulation. This gets complicated if the new micro-simulation gets out of sync with the orginal and in this case the system may have to switch other regions from record mode to full simulation. The point is that it is easier to handle this mixture of recorded and actual simulations using virtual machines than having to control it all "by hand".
The discussion that ended the workshop brought all these points together. Virtualisation makes grid or cloud computing more feasible, which may make HPC resources available to more people. New CPUs have explicit support for virtualisation that will make it almost cost-free for the data centre, and there is a good chance that these advances will also reduce the overhead for some HPC applications too. The key problem seems to be predicting which applications will run smoothly on a virtualised system and which will suffer noticeable slowdowns. As always, more research is needed!
This was the topic of yesterday's workshop on System Level Virtualisation for HPC (which was part of EuroSys 2008). The workshop was rather small but I did learn quite a bit.
A couple of talks investigated the performance hits of using virtualised OS's instead of addressing the hardware directly. This varied, depending on the type of application; if IO was minimal, the slowdown was minimal too. An "average" slowdown seemed to be on the order of 8%.
Samuel Thibault of Xensource looked at ways of using the flexibility of a hypervisor to implement just those parts of an operating system that are absolutely necessary - ignoring more general-purpose facilities such as virtual memory, pre-emptive threads and the like. Essentially, this was using Xen as a "micro-kernel done right".
Larry Rudolph, who works for VMWare but was here speaking with his MIT hat, gave a scenario where virtualisation might benefit the HPC user. Some problems require massive machines that researchers only get access to maybe twice a year. The example Larry gave was modelling global ocean currents. If the researchers can save the messages sent between regions of this model, they can later run more detailed analyses on one particular region on a much smaller machine, using the recorded messages to playback the effect of the rest of the simulation. This gets complicated if the new micro-simulation gets out of sync with the orginal and in this case the system may have to switch other regions from record mode to full simulation. The point is that it is easier to handle this mixture of recorded and actual simulations using virtual machines than having to control it all "by hand".
The discussion that ended the workshop brought all these points together. Virtualisation makes grid or cloud computing more feasible, which may make HPC resources available to more people. New CPUs have explicit support for virtualisation that will make it almost cost-free for the data centre, and there is a good chance that these advances will also reduce the overhead for some HPC applications too. The key problem seems to be predicting which applications will run smoothly on a virtualised system and which will suffer noticeable slowdowns. As always, more research is needed!
Comments