At SC’10 the most interesting concept that came to my attention was not from the show floor, nor any formal meetings, but instead came from a dinner I attended Wednesday night before I left at Brussard’s. Most of these ideas were not originated by me and have in fact been around for several years. I’m just catching on. The concept is what I think will be the next disruptive technology in high performance computing. It has to do with how server vendors can use micro servers in their product lines.
Consider the following:
– Multi-core Server processors sell for a larger premium than the processors that are found in desktops, laptops, and iPads. These commodity processors are becoming more powerful all the time. I am still extremely pleased with the performance of my laptop, mobile phone, and desktop, even though I bought these last year.
– Just as commodity Intel Xeon processors took over more expensive RISC based processors in HPC applications and now dominate the top500 list, could it be possible that commodity processors are getting “good enough” to make a compelling compute block? Could it be that clustering multiple commodity processors together could be more cost effective in some cases than today’s server processors?
– Today’s multi-core servers have more horse power than a normal OS or workload requires. Virtualization is used to partition hardware components so that multiple operating systems can use it without wasting resources. (Consider the uptime report that showed that 70% of servers in the data center are idle). The main problem with Virtualization? Decreased I/O performance
So to summarize these points:
- commodity processors found in desktops may be more cost effective than server x86 processors and are getting better in terms of performance.
- virtualization is weaker in I/O performance than physical servers. (Yes, I know there are ways to optimize it, but take apples to apples here)
Now take a look at http://plugcomputer.org and look at the Sheva Plug. You can buy it for $99. Now think how many of these you could pack into a datacenter. Combine this with virtualization software from ScaleMP or other virtualization software and you have the most flexible, cost efficient, data center ever.
Or imagine this: What would be more cost effective? A 4U IBM x3950M3 with 4 sockets and 96 DIMM slots or a future system that takes the same amount of space but crams perhaps 64 of these micro servers inside of it?
Naturally, this comes down to management. If its just as easy to manage physical machines as it is virtual machines, then this technology will take off. Managing physical servers at scale has always been a challenge to most IT shops. This has been our passion at Sumavi and with xCAT. If we could make it just as easy to deploy these servers as it is to manage virtualized servers, then the data center would take a dramatic shift just as it did when all those x86 servers started showing up and replacing old Unix RISC processors.
But there are challenges:
- I/O: If all these micro servers were plugged into some type of array of servers getting data out of them could prove tricky. I am excited to see what type of solutions server providers come up to address this. Will they create a new blade form factor? Will they shove them into the box like disk arrays? Will they use InfiniBand technology or some type of proprietary interconnect?
- How will virtualization be incorporated into this? I think the future looks to be a healthy mix of virtualization with physicalization. Let’s face it: even these commodity processors have more horse power than some apps need.
- Mass provisioning of operating systems. Management is what I see as the biggest game changer. Make this easy, and you’ve got a gold mine. I think this is where xCAT could play a major role. In addition, going stateless will increase agility and lessen deployment times to the same amount as a VM.
There is lots to read up about this exciting space in physcialization:
ARS Technica has written some interesting articles. While I don’t look at physicalization as a replacement for virtualization, it certainly has a place.
Rackable has had an offering. I’d be interested to learn what SGI is doing with that now.
Let me know if there are other links, articles, or products that we should look into!