HPC is often used as a general term but it is actually many different facets depending on the computing model.
CERN is at the centre of a server grid of 100,000s of servers called WLCG (http://wlcg.web.cern.ch) for analyzing the data from the Large Hadron Collider. The servers are located at over 200 sites worldwide in a tiered structure. However, we're more of High Throughput Computing (HTC) rather than HPC. HTC has a large number of programs running at the same time which have no need to talk to each other. Thus, it is more like a large scale batch farm than a massively parallel machine. While we lose a little performance on memory and I/O performance, virtualization brings major benefits in ease of management of the thousands of servers. We expect to be saving the few percent of overhead over the lifetime of machines by more flexibility scheduling repairs and placing the workload, such as overcommitting a hypervisor when one of the VMs is waiting for a tape to be mounted. Analysing the 25PB/year for the next 20 years, we're pretty intensive compute and I/O load. However, when we take the total cost of ownership, people included, we expect a significant efficiency gain from the use of a private cloud. Some more details at http://cern.ch/go/NH9w The massively parallel processing use cases with Crays/BlueGenes may not benefit from private clouds but many of the research sites will. Tim On 3 Jul 2012, at 16:12, Matt Joyce wrote: On Tue, Jul 3, 2012 at 2:01 AM, Simon G. <semy...@gmail.com<mailto:semy...@gmail.com>> wrote: Secondly, I don't think we shouldn't compare GCE to Openstack. I understand that right now cloud (Openstack, Amazon, ...) is just easy in use, managed and scalable datacenter. It allows users to create VMs, upload their images, easily increase their (limited) demands, but don't you think that HPC is the right direction? I've always thought that final cloud's goal is to provide easy in use HPC infrastructure. Where users could do what they can do right now in the clouds (Amazon, Openstack), but also could do what they couldn't do in typical datacenter. They should run instance, run compute-heavy software and if they need more resources, they just add them. if cloud is unable to provide necessary resources, they should move their app to bigger cloud and do what they need. Openstack should be prepared for such large deployment. It should also be prepared for HPC use cases. Or if it's not prepared yet, it should be Openstack's goal. HPC in the cloud operates more like a grid computing solution. With things like Amazon HPC or HPC under openstack the idea is to allocate entire physical systems to a user on the fly. Traditionally to date that has been done with m1.full style instances. In many ways bare metal provisioning is a better option here than a hypervisor. And for many people who do work in an HPC environment bare metal really is the only solution that makes sense. The reality is that HPC use cases lose a lot of the underlying benefits of cloud infrastructure. So they really are something of an edge case at the moment. I believe that bare metal provisioning from within openstack could be a bit of a game changer in HPC, and that it could be useful in a wide variety of areas. But, ultimately I believe the usage that HPC in no way reflects general computing needs. And that really sums it up. Most folks do not need or want HPC. Most folks with HPC needs don't want a hypervisor slowing down their memory access. I know that clouds are fulfilling current needs for scalable datacenter, but it should also fulfill future needs. Apps are faster and faster. More often they do image processing, voice recognition, data mining and it should be clouds' goal to provide an easy way to create such advanced apps, not just simple web server which could be scaled up, by adding few VMs and load balancer to redirect requests. Infrastructure should be prepared even for such large deployment like that in google. It should also be optimized and support heavy computations. In the future it should be as efficient as grids (or almost as efficient), because ease of use has already been achieved. If, right now, it's easy to deploy VM into the cloud, the next step should be to optimize infrastructure to increase performance. Apps are actually slower and slower. The hardware is faster. The Applications themselves abstract more and more and thus slow down. As for what you do on your instances, that's entirely your own thing herr user. Some large data and some serious compute use cases simply don't lend themselves to cloud today. Hypervisors are limiting in so far as they give up some speed to provide the ability to share resources better. If you have no desire to share resources then virt machines become something of an impediment to you. So I don't see this as being accurate for some use cases. There are also other external limiting factors. People don't just turn on a dime. Many of the scientific and industrial applications of computing power are built around software stacks that have grown over time, and for a long time. Those stacks can't be made to easily adopt the benefits of a new technology. Sometimes the reason not to use cloud as a platform is entirely related to your inability to modify an existing software suite enough to make it worthwhile. I have seen this before at super computing facilities. I've always thought about clouds in that way. Maybe I was wrong. Maybe cloud should do only wh)at it's doing right now and let to others technologies handle HPC. I think many, in the HPC environment, argue this is probably true. I don't necessarily agree. GCE obviously proves a point. Sharing resources means that you don't have to run your own super computer. You can simply rent enough of a compute environment to solve your problem at will. And odds are the environment will be pretty up to date. For many use cases cloud environments are just dandy. And HPC offerings in IaaS providers are getting better all the time. For low funded research, citizen science, and a million other small fries out there there is certainly a value in lowering the barrier to entry in this technology. That being said, I think that private HPC will never go away if only because of data retention rules and law. Much research deals with data that must be either safeguarded or simply classified and placed in an environment that meets that classification levels needs. In some cases those limiting factors can make working with the amazons or googles or rackspaces of the world an impossibility. So on one hand, yes I think HPC in openstack is important. And will grow still more so as time goes on. But, on the flip side I believe HPC user requirements do not reflect the needs of general computing users. From a technical backend perspective the likelihood is most businesses have no necessary use of HPC. And bursting to a public offering probably makes a lot more sense than on their own private pond of compute resources. One future environment might see an openstack environment local to the org that has users testing instances and prepping them then sending them out for a few days at a time every few months to crunch some data set on an HPC environment. In that case the openstack environment would become an HPC instance proving ground / staging area. But hey, maybe I am wrong. =D -Matt _______________________________________________ Mailing list: https://launchpad.net/~openstack Post to : openstack@lists.launchpad.net<mailto:openstack@lists.launchpad.net> Unsubscribe : https://launchpad.net/~openstack More help : https://help.launchpad.net/ListHelp
_______________________________________________ Mailing list: https://launchpad.net/~openstack Post to : openstack@lists.launchpad.net Unsubscribe : https://launchpad.net/~openstack More help : https://help.launchpad.net/ListHelp
