The HPC Cloud






What is HPC (High Performance Computing)? It is something big and expensive, hard to use and administrate, that consumes a lot of power, can be used only by an elite group of HPC specialists. No two HPC sites are similar, and mostly governments, non-profit agencies and the like are giving money. Traditionally - there is NO expectation of $ profits. HPC is sort of sacrificial price society pays to advance science in a political sense. The only measurable attribute to define success was the participation in the TOP-500 list. Using a 25-year old performance test, called LINPACK, the biggest goal is to rank any new HPC site among the top 500 according to some metrics called Rmax and Rpeak that mean absolutely nothing to the 99.99% of the population.

An HPC site is a cost center. It gobbles and gobbles money, in return for a favorable Rmax ranking

A large IT site uses 150 MW of energy and costs about $60 million to built. (see Horst Simon, Lawrence Berkeley National Lab, slide 20)

The electrical utility in California, - where I live - is named PG&E. Last month, quietly, they increased the rate for residential power to 24 cents per KW from 22 cents. At this rate, the HPC site will have to pay $36,000 per hour or, for full 365 days. As far as I know each day has 24 active computing hours, we land at $315M per year in energy costs alone. Even assuming a 90% discounted industrial electricity rate versus residential rates, this is still whooping $2.5 millions per month or $30M per year.

This is equivalent to employing 200 highly skilled engineers

Sure we can not have HPC installations like this in NYC, Sacramento or San Francisco itself. PG&E is an HPC show stopper in California, which ironically is the state where Silicon Valley is and where the great innovator companies have head offices.


Here is how HPC will be saved and becomes main stream: we will build a cloud named HPC-cloud that meets what the users seek. I mean any user, the type of user who uses Google today, will be able access a directly or indirectly an HPC-cloud. That significantly much larger number of HPC users (in comparison to the HPC elite today)

- will always have all resources s/he needs
- will pay only for what it uses
- will use the HPC applications as a service

The users will have no idea where the HPC-cloud is located and what's inside the HPC cloud. They are shielded from all complexity. Power Optimization software will turn off the power on all unused resources when demand is low. Turning the power on means that there are enough paying users to cover the cost of the additional power from the billings sent to users.

An HPC-cloud will not be a cost center. It will be a Profit Center. This means it can be sold to the commercial enterprises. The sale is much simple. Companies that need HPC (and today with all risk analysis and security concerns most companies do need HPC), will loose money every single day that did not buy an HPC cloud for their employees and and their customers.

I believe in the Adam Smith principle promulgated in 1776: "Individuals trading freely with one another, following their self-interest leads to a growing and stable economy." It is time to free the HPC from eternal subsidies from governments and military, or iron-fist bureaucratic regulators.

It is the time to democratize the HPC, by making it a business that meets real needs for all the people engaged in producing wealth - in USA and elsewhere in the world.

Comments

Srini Chari said…
I think HPC is a broad field...of which one segment could be the large government and academic institutions. These could be cost centers...as these organizations are not traditionally charged with making profits. But there are are several industrial applications of HPC - financial analysis, Computer Aided Engineering, Seismic Analysis, etc..where HPC is a crucial part of a business process needed for competitive reasons or for better products/services. The increased use of HPC for industrial uses is one of the key drivers of HPC in a Cloud Model....It's really a business reason not a technical driver.
Jay said…
This is already being done by companies like Tata CRL and Sun's network.com
Anonymous said…
I agree with Srini. HPC is not just large government sites with very large machines. HPC is in many companies and organizations. People assume that HPC means 'really large machine, running big parallel jobs' but the majority of 'HPC' is really lots and lots of throughput computing on large clusters of 'mostly Linux machines'. These customers with 'large throughput' clusters are the first to use cloud computing, not the large parallel clusters. I honestly don't see large HPC going away, but I do see a dramatic growth in medium sized throughput clusters. These clusters as Srini points out are a 'crucial part of the business process'.

Popular Posts