04.05.07
Posted in Google NYC at 7:30 pm by soumitra
Maximizing usage of available power capacity is the key!
Power subscription at the facility level seems promising (whatever that means)
large systems should utilise power efficiently.
Q and A going on…
… almost ending, so i guess i can switch this off.
Oh
I forgot.
Conclusion (from Luiz’s slides):
Power/ energy efficient and fault tolerance are central to the design of large scale computing systems today.
Technology trends are likely to make them even more relevent in the future, increasingly affecting smaller scale systems.
Permalink
Posted in Google NYC at 7:13 pm by soumitra
Power efficient datacenters:
Idle machines also consume power.
Compared to humans, power consumption for machines is waay high, since humans use less power while resting, but high while performing. Machines don’t have that high gap between idle and high activity power consumptions.
Permalink
Posted in Google NYC at 6:59 pm by soumitra
“Fault Tolerance systems lie.
Disk failures: Failure rates < 1% per year.
Study conducted proves that irrespective of age, disk failed... (couldn't catch that)
also, Temperature is not a major criterion for disk failures.
SMART disks…self monitoring. (!)
Even then, 56% of disks failing had no predictable behavior.
Well, more bad news: 70% of drives survive for 8 months. (So after 8 months, no predictability). So, SMART is not a great solution.”
Permalink
Posted in Google NYC at 6:44 pm by soumitra
So on stage, Luiz Barroso!
“Single fault will bring down the large scale computer systems that use shared memory programming model, and fault containment is complex. A difficult problem”
Permalink
Posted in Google NYC at 6:44 pm by soumitra
“The MHz race (Moore’s law) just makes the machines faster… the programs coded 10 yrs ago automatically run faster…
Additionally, power consumption increases and temperature management is a concern.
Ergo, cost increases with time.
So now, the equation becomes higher computing power = higher power bills= higher costs”
Permalink
Posted in Google NYC at 6:43 pm by soumitra
“Power supply units of machines are less efficient. They can be made up to 90% efficient, though they are more expensive.
Multi-core processors solve the problem a bit, but there should be enough threads.”
Permalink
Posted in Google NYC at 6:30 pm by soumitra
“Crosstalk at CMOS level, considering their shrinking size. Disk drive failures are increasing…”
Permalink
Posted in Google NYC at 6:20 pm by soumitra
Free beer and wine, and supposedly delicious hors d’oeuvres… ( i didn’t try any …yet!)
Mike here is giving out deets on the next in the Speaker series..
Permalink
Posted in Google NYC at 6:18 pm by soumitra
Am sitting here at the Google’s NYC office… they have a Speaker series and i got myself invited.
So we’re all gathered here, and all guys are like filling in slowly. the talk was supposed to start at 7ish, but then the rush is so much that they’re still settling down.
Okay…just starting…
Permalink
