Richard Cownie (tich@pobox.com) on 9/17/09 wrote:
---------------------------
>anon (anon@anon.com) on 9/17/09 wrote:
>---------------------------
>
>>>I think Google have figured out the scaling so that
>>>absolute performance of the nodes is not so critical,
>>>within reason ...
>>
>>Why did you cut out the arguments I made as to the opposite?
>
>Well, my recollection was that Google were using small
>cheap machines. But actually, looking at the article again
>it does seem that they're dual-socket machines stuffed
>with DIMMs, so maybe my impression was wrong. Anyway,
>you can read the article and judge for yourself.

I judge based more on what I've deduced for myself (ie. footprint per unit of CPU, and workload quantization / fragmentation effects).

I know they run several workloads together on each system, using containers and resource management in the Linux kernel, and they try to fit the workloads together to make maxiumum utilization of each major component. I also know they have some SMP scalability problems in some areas of the kernel, which suggests at least some systems have a reasonable number of cores.