3
$\begingroup$

Following the latest trends in linear algebra, power consumption is quickly becoming an important factor to consider when designing and using an algorithm.

I've even heard that over time, the electricity bill can exceed the initial purchasing cost of a supercomputer and people predict that in the future, HPC resources will not be priced in terms of how many compute nodes you used for what amount of time, but (at least partially) in terms of how much power was consumed during the calculations.

My question is now: Does anyone know if this has already happened? Are there any supercomputing centres that charge for consumed power?

$\endgroup$
4
$\begingroup$

To my knowledge, yes and no.

When HPC centres (many of them) calculate the bill for companies and private individuals, they assume some average amount of power is going to be used and charge accordingly. Also currently ComputeCanada determines the value of allocated resource grant by taking average and max power consumption into the account. I know that this is also the case with PRACE, but I do not know the policies RIKEN of Japan, XSEDE of USA or other national laboratories in USA. However, this approach distributes the cost across all users and creates almost a tragedy of commons scenario. Some researchers tend to write terribly inefficient code and use incredible amounts of computing power in order to gain speedup. So, recent trend is to individualize the cost so people do not write inefficient codes and waste the resources. Not only the bills will be individualized, but the researcher will be expected to show efficiency of their codes in their grant application and provide grant reports related to their use of resources.

However, in general, policy is a fickle thing and by its nature, political. For example, there is a great deal of opposition from researchers who want to keep their research "life style". Their immediate focus is not high performance computing, they just use it as a tool and they do not want to spend time optimizing their codes. Because such a practice has no publication value for them. On the other side of the aisle, there are people concerned about wasting tax-payers' money, global warming and electronic waste and they argue that more regulatory oversight is required. One primary example I can think of is the idea of employing researchers as code developers and maintainers at the HPC centres to improve longevity and efficiency of existing codes in addition to support development of new high quality and high performance scientific software.

Note: Top500 is also tracking Green500, so there is some momentum towards making clusters more efficient too.

$\endgroup$
2
  • 1
    $\begingroup$ At least in the US and some supercomputing centers in Europe, applications for supercomputer use have had to show efficiency of the code (e.g., how well it scales) for many years. $\endgroup$ Aug 27 '20 at 2:57
  • $\begingroup$ I don't know much about how US institutions allocate their resources in general. However, for the few cases I am aware of, the researchers can write grant proposals for provisional resources to develop and test their codes and these resources are up to 256 processes or sometimes 512 processes. I think this is mostly practised at national laboratories, though. Of course, access to 1000s or more processes are locked behind diligent revisions - which is as it should be. $\endgroup$ Aug 27 '20 at 5:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.