Usenix ATC 09 Awards & Keynotes

Jun 18, 2009 Conferences

Best Paper Awards

The first best paper award went to Grzegorz Miłoś for Satori: Enlightened Page Sharing. I’m a big fan of memory sharing between virtual machines, so I’m glad to see some recognition for this type of work. I talked with Irfan Ahmad from VMware after the talk and I have to agree with his view that the real benefit of this type of system is not in attempting to free up memory for other VMs, but in reducing I/O latency since fewer blocks need to be read from disk.

Next up was Tolerating File-System Mistakes with EnvyFS, which I haven’t read yet, but now I’ll have to take a look.

Keynote

The keynote was by James Hamilton from Amazon Web Services. He made some pretty interesting points about how enterprise costs are so different from services costs. I’m sure that in time, enterprises will do the best they can to get closer to the services model by eliminating their “people” costs (sorry folks!) and trying to make larger scale homogenous systems. The talk did a great job at providing the high level view for where the real problems are in big data centers. I also recommend his blog for anyone not already checking it out, it is full of a wealth of interesting data (plus some good ideas). The slides from his talk are available on this page.

• Enterprise’s main cost is people
  • Often about 100 servers per admin
  • Have many different apps, each with relatively small scale -> difficult to automate
• Services world’s main cost is hardware
  • >1,000:1 server:admin
  • Don’t look at raw performance, look at work done per dollar, or work per joule.
• Data Center monthly costs (3 year amortization for servers, 15 year for infrastructure)
  • 15MW data center ~$200M
  • Servers 50%
  • Power & Cooling infrastructure 25%
  • Power 22%
  • Other infrastructure 3%
  • Even combined, power and power infrastructure is less than 50%
  • but server costs are decreasing, while energy is not…
  • So current headlines are wrong, but still correct
  • Take away: server cost is still very high, so it makes sense to USE servers if at all possible. Turning them off only saves on the energy costs, which is relatively small (22%)
• PUE = (total facility power / IT Equipment power
  • 1.7 PUE is a “normal” new Data center (wastes 0.7 watts per 1 watt used by servers)
  • But PUE can be deceiving since it counts things like server fans as useful energy
  • tPUE is new metric that just counts useful server energy – see blog post for more info
  • Key points:
    • Server efficiency and utilization is a good thing to improve
    • Cooling waste is unreasonably high
    • Power distribution waste isn’t too bad
    • When provisioning energy, don’t assume DC will have all servers running at peak load. “Oversell” power, and then shed load to other data centers if somehow all servers simultaneously ramp up to full load.
• Temperature
  • Most people run at 81 degrees, but systems can handle much higher (Dell 95, Rackable 104)
  • Raising temp within the data center can save lots of money (especially if it is cool outside)
• Resource Consumption Shaping
  • Apply resource optimization across entire data center
  • Move work from peaks into valleys since costs are based on peaks

My only disappointment from the keynote? His hair wasn’t as big as I’d imagined