Virtualization and the Grid
I just got pinged by a friend who I’d recently discussed virtualization with. He asked me to blog the relationship between virtualization and grid computing. It won’t take long so I may as well:
A technician from one of the major hardware vendors told me that only about 60% of the currently deployed server population could be virtualized. The reason for this is simple:
You cannot share a saturated resource.
You can beef up the memory on a server (or blade) to some degree, but you cannot do much about i/o when it’s saturated or cpu when it’s saturated. So the servers that won’t virtualize are the ones that manage heavy workloads that max out on cpu, i/o or memory.
Many of those servers operate in SMP configurations running very high workloads. I should think that the majority of Google’s hundreds of thousands of servers are not virtualized and never will be, because Google runs a very large grid where every resource is made to work for its money. The same is probably true of Amazon, Yahoo, MySpace, etc. Virtualization has no relevance to such large server configurations.
So we can view virtualization as “the opposite” of a grid. With a grid we combine many blades or servers to manage large workloads (like Google does for search or like Oracle did with RAC for data warehouses). Ultimately, that’s why only 60 percent of servers will ever be virtualized.
The Tipping Point
If we move to a heavily virtualized world, the question arises as to what the basic computing unit should be. Should it have a cpu with 2 cores, 4 cores, 64 cores or what. Maybe it should have several. How much memory should it cater for; 16 gb, or 32 gb or 64 gb or 1024 gb? and how much i/o should it allow for?
The hardware vendors are going to have to scratch their heads over this one for a while. On the client side Intel and AMD can go on playing “double the power” and maybe there will be a market. But on the server side it’s all “back to the future” with the server vendors needing to manage resources efficiently – like a mainframe. Some of that means virtualization and some of it means grid.
This is a posting in the Virtualization Focus Series. Click here to see an index of such postings.



















