| 1 |
On 24/06/2015 13:50, Alec Ten Harmsel wrote: |
| 2 |
> P.P.S. Also, on 1% better performance: My professor for the compilers |
| 3 |
> class I took used to (maybe still does) work at Google. Apparently |
| 4 |
> Google sees a <1% increase in performance as *the best thing ever*, |
| 5 |
> because it can save them a bunch of money in infrastructure and power. |
| 6 |
> Apparently Google are the ultimate ricers. |
| 7 |
|
| 8 |
|
| 9 |
Sounds like a case where Google already did the sensible optimizations |
| 10 |
long long ago and are now hitting the diminishing returns from the long |
| 11 |
tail. There are probably many of these and they all add up. |
| 12 |
|
| 13 |
One thing I've learned about Google's setup - there's nothing else like |
| 14 |
it out there and they are truly unique. Almost nothing Google does to |
| 15 |
optimize their setup is widely applicable to anything else :-) |
| 16 |
|
| 17 |
Take their power density. Last figures I have is they were running at 4x |
| 18 |
the kW per square foot as anyone else with a brain. This terrifies |
| 19 |
people who know about cooling. But, that's the setup and that's what |
| 20 |
Google has to work with. Now suddenly, all those lots of little |
| 21 |
improvements start to become a huge deal. |
| 22 |
|
| 23 |
So yes, ultimate ricers. Also the ultimates in |
| 24 |
"riding-co-close-to-the-edge-you-fall-off-the-cliff" :-) |
| 25 |
|
| 26 |
-- |
| 27 |
Alan McKinnon |
| 28 |
alan.mckinnon@×××××.com |
Archives