CPU resources scale beyond one server?

Portal Home > Knowledgebase > Articles Database > CPU resources scale beyond one server?

Posted by Eiv, 06-24-2011, 03:15 AM
Hi WHT I am relatively new to cloud. I am just wondering which cloud platform allows you to allocate multiple server's CPU to a one single customer. OnAPP cant do this right?
Posted by nix101, 06-24-2011, 03:32 AM
OnApp doesn't allow it. A load balancer + multiple cloud instances should work by distributing instances across multiple hypervisors.
Posted by boskone, 06-24-2011, 03:33 AM
As far as I am aware there is no current cloud/virtualisation tech that alows you to share the CPU from multiple physical servers to a single virtual machine.
Posted by Eiv, 06-24-2011, 03:36 AM
What about Applogic and Amazon?
Posted by boskone, 06-24-2011, 03:37 AM
Nope. It isnt technically possible at the moment. It's another misconception of 'cloud' that your server / service will magically straddle multiple servers (physical) As above, the only option is use Load Balancing to spread the work across a lager number of smaller machines, which adds up to the same end result...
Posted by Eiv, 06-24-2011, 03:44 AM
So much for the cloud hype. Then it is pretty much like VPS technology except you get better redundancy.
Posted by boskone, 06-24-2011, 03:46 AM
It's 'better' VPS technology with all the control, redundancy, failover, scalability and management of a dedicated cluster. It just not magic
Posted by eming, 06-24-2011, 04:46 AM
There are no hypervisor technology that can do it so far. It's not an OnApp thing, but a hypervisor limitation. The way OnApp sort of works around it is by using our horizontal autoscaling technology. Basically if your hypervisor is out of resources, OnApp will make a clone of it and loadbalance the incoming traffic between the two (or more) nodes automatically...all in realtime and with no manual intervention. I agree, the 'cloud' term is not doing us any favours. I was hoping for it to go away soon, but not sure it will happen (damn you, iCloud). D
Posted by dazmanultra, 06-24-2011, 05:38 AM
That is true if you're looking at the Infrastructure as a Service (IaaS) type systems - these are tools to allow you to build your own systems based upon the cloud infrastructure. They provide building blocks for you to use. The Platform As A Service (PaaS) hosting options out there are the ones that are designed to leverage this type of infrastructure for you, without you having to worry about load balancing or provisioning new web server instances.
Posted by Krazy, 06-24-2011, 06:40 AM
was under the impression that , cloud can provide collective power, damn! the biggest possible system is limited to the individual server capacity then
Posted by JasonD10, 06-24-2011, 10:29 AM
There's a lot of benefits of Cloud over VPS such as horizontal scaling, and ease of vertical scaling. There is redundancy in the environment, and ease of management for a provider as we manage Clouds are one large bulk of resources, instead of a bunch of individual VPS servers working alone. Uptime, performance, and consistency are some direct end-user results and improvements over VPS. Scaling beyond one physical server I would say is definitely easier with some platforms than others. Like you mentioned, AppLogic being one of them. It's inherently built in the infrastructure to have fully integrated clusters providing the capability to vertically scale beyond one server with mini "appliances" that are single purposed. The level of integration is a big thing with AppLogic to not be overlooked. Think of it like building a traditional physical server cluster spread out all throughout a Data Center, or having a private cage with direct and private switches, all servers directly connected to each other, shared storage between them, the ability to quickly and effectively migrate components between each other, configure groups so the cluster servers don't run on the same physical servers for high availability (without pinning them to specific servers), prioritizing IO loads, and so on. AppLogic is a beast of a platform to build Clusters from, and mimics traditional non-Cloud server clusters as best as we've seen. So if you have a large scale application that requires the vertical scalability beyond that of a single server, I'd highly recommend focusing on providers that use AppLogic.
Posted by DeanoC, 06-26-2011, 02:38 PM
It does exist, however its still at the experimental stage. Its known as a Single System Image (SSI) cluster where a custom kernel provides a view of the entire cluster as a single entity. Its pretty amazing when it works (/proc/cpustat looks cool! ) If you have a genuine need for it versus the more traditional scaling technologies, get someone whose done it but tbh you probably don't need it.
Posted by kris1351, 06-26-2011, 11:06 PM
It's coming I am sure, but don't expect it not to include a price tag.
Posted by UNIXy, 06-26-2011, 11:56 PM
The technology has existed for quite a while. It's called NUMA. But it's too expensive and impractical for the Web hosting segment. Impractical because of the latency involved in fetching data across physical nodes. Load balancing is a cheaper workaround. IBM had shipped the x440 P4 Xeon server line with proprietary expansion cabling technology that allows chaining a few x440s together. You can then leverage the CPU and disk IO processing of the resulting chain (/proc/cpuinfo lists all CPUs from all banks & /proc/meminfo recognizes memory from all banks / chassis). Opteron CPUs (and then later Nehalem) have started implementing NUMA but only as a performance boost on multi-core / SMP nodes. You'd need some kind of chassis interconnect to chain up nodes. Regards
Posted by kris1351, 06-27-2011, 09:33 AM
Sun has offered it with blade based Enterprise servers for years, we did this with the E10ks a long time ago but cost/space/power were huge. Just the chassis of a E10k was over 200k in price. It's out there, but on a hosting level I suspect we will see it come at some point as we have seen the elastic clouds.
Posted by iTom, 06-27-2011, 12:59 PM
I wouldn't call this a software/hypervisor limitation - the back plane speeds required between physical servers just doesn't exist yet, not even on a blade server. Your going to be much better using multiple machines with a load balancers, if that's not possible the application your running will need to be optimize or potentially redesigned.
Posted by DeanoC, 06-27-2011, 01:31 PM
Only shared memory architectures need extremely fast memory links. SSI is usually done at the fork/thread level, whilst you still need a fast network you don't need extreme communication bandwidth in some cases.
Posted by tchen, 06-27-2011, 09:00 PM
Shared anything networks are just horrible to scale. Unless you use some sort of MPI between your forks and threads then you're just running into distributed lock issues. And if you're using MPI, why are you sharing anything? Really, the interest-set for any piece of shared data needs to be controlled and quite small for you to horizontally scale effectively. It's not a trivial problem given physical limitations and network effects and there are many solutions out there, each applicable to certain types of problems. There's really nothing singular that would fit everyone's needs.
Posted by tulix, 06-28-2011, 01:37 AM
Sorry, but I am surprised that nobody has mentioned 3TERA - a cloud software which DOES allow to scale beyond one server and create a super computer (it is based on a grid technology - not a new stuff)
Posted by tchen, 06-28-2011, 06:08 PM
CloudWeb already mentioned AppLogic. Caveats apply as usual regarding the type of applications that scale on it.

Add to Favourites Print this Article