Slashdot Mirror
Benchmarking Power-Efficient Servers
modapi writes "According to the EPA, data centers — not including Google et al. — are on track to double power consumption in the next five years, to 3% of the US energy budget. That is a lot of expensive power. Can we cut the power requirement? We could, if we had a reliable way to benchmark power consumption across architectures. Which is what JouleSort: A Balanced Energy-Efficiency Benchmark (PDF), by a team from HP and Stanford, tries to do. StorageMojo summarizes the key findings of the paper and contrasts it with the recent Google paper, Power Provisioning for a Warehouse-sized Computer (PDF). The HP/Stanford authors use the benchmark to design a power-efficient server — with a mobile processor and lots of I/O — and to consider the role of software, RAM, and power supplies in power consumption."
Everyone is looking at virtualization for this sort of thing, and it does hold some promise. Currently, though, virtualization still comes with some very significant performance penalties. I think if virtualization can further mature, and we can get more cores and cheaper memory, we will be where we want to be.
Eventually, if RAM continues to get bigger and cheaper, more cores get packed into chips, and virtualization becomes what it is intended to be in terms of performance and stability, we will start to see a lot of movement away from lots of small pieces of commodity hardware and back to fewer large systems in the data center. Of course, cheaper and faster disks on Enterprise storage systems like Netapp or EMC will help that transition too.
Years ago we heard how PCs were going to be embedded in everything from the dishwasher to the refrigerator, and I was left wondering, "why?"
Perhaps now I know.
It would be nice if I could set my house up on a "power budget", and let my appliances vie for electrical power and load-balance themselves to stay within that budget. If all appliances spoke over the in-house wiring (or perhaps wireless) and could turn themselves off or adjust their power usage that would be awesome.
You could implement something similar to this today with an X10 system or the like, but this is more of an off/on scenario, and is not based on actual power demands.
It would be great if all of my electrical things in my house could get together and say, "OK, guys, we have X amount of electricity to use today between all of us. Let's figure out, based on past usage patterns, who needs to be on and when in order to hit this budget".
A work that expires before its copyright never enters the public domain and thus enjoys eternal copyright protection.
Answer me this : how much power is lost through the use of inefficient programming languages and architectures which only emphasize processor speed, instead of balancing memory, processor and IO ?
Python, Perl and PHP all suffer from one big drawback : when you scale up you need that much extra processor power. One programming language I know (Common Lisp) offers the advantages of them, but can be compiled to near C/C++ speeds. I suppose there are others. Don't come saying that programmers are expensive. It seems that what you gain on programmers, you lose in the cost of your datacenter. I don't know how Java matches here, it probably depends upon the deployment of more recent JIT compilers.
If you see how much a process has to wait on IO, how come there are still no good solutions in providing enough IO bandwidth that the processor can use fully ? (Unless you buy a mainframe or iSeries system that is)
Just asking.
Just junk food for thought...