Pat Helland gave an awesome keynote last week at the p&p Summit. I wasn’t able to blog about it until today as he gave the same talk at TechEd in Europe yesterday. You can read the abstract here. He got through over sixty slides in an hour so this just covers some of the things that really grabbed me, and a lot of his talk really grabbed me. I’ve been thinking a bit about Green IT since building a low power home server a couple of months back and Pat gave me a load more stuff to think about.
Power is a huge issue
Energy costs more than the hardware so it’s cost effective to update inefficient servers.
Power Usage Effectiveness, PUE = (Total power) / (IT equipment power)
|Less redundant DC||1.2|
PUE and infrastructure cost rises dramatically as redundancy requirements increase. Redundancy amounts to 20% of the datacenter costs
What does this mean to developers?
- Virtual machines – allows more flexible DC resource allocation.
- Looser SLAs – High level SLAs require hardware with lower utilization to meet the SLA. Lowering the SLA allows for higher hardware utilization.
- Shallower calls – deep call stacks place higher SLA demands on the services at the bottom of the stack. Shallower call stacks allow for looser SLAs. Typical Amazon page request hits about 150 servers placing big SLA requirements on the data layer or the business needs to build eventually consistent systems.
- Read-only data – cached data on lower availability (cheaper) data centers
- Eventually consistent
- Queuing – very long SLAs
My key takeaways from Pat’s talk were:
- Power is a huge issue for servers and datacenters:
- 2.0% of total US electricity, doubling every five years (source EPA).
- Infrastructure and servers are driving enhancements for energy:
- Power distribution and cooling are being improved.
- More efficient servers.
- When things start to cost real money people take notice and try to make savings.
- Current power costs mean that it’s cost effective to replace inefficient servers in the DC.
- Today’s tragedies:
- Datacenters tend to over build and under use, they also don’t share between different datacenter companies.
- Tragedy of the commons: Overuse of shared power and cooling and over dependence on high availability servers.
- Tragedy of the anti-commons: Isolation of servers and infrastructure per service.
- Software is demanding datacenters that make these tragedies inevitable
Wow… That’s a lot of stuff and only about 1% of what Pat had to say. If you’re at TechEd EMEA in Barcelona hopefully you caught what Pat to say in person.