[ ... ]
> So, "disk" means different things in different contexts.
> Graeme's presentation uses "disk" to mean "storage in an SE".
>
> Be careful when using the term "disk bound" in this context, therefore
> - T2s are limited in the number of jobs they can run by the number of
> CPUs (which is what Graeme's commenting on), but the performance of
> any particular analysis job on a T2 node
Ah yes, indeed I really meant "disk capacity bound" in that
context (archival storage). That is it seems that T2s should buy
more CPUs and T1s should buy more storage.
> is almost certainly IO bound, not CPU bound.
And I suspect that EwanMM was asking "how much" so he can make
tradeoffs as to power consumption/cost (I guess).
It would be nice to have some guidelines from other experiments
to, and for production as well (but see below).
> [ ... ] performance of the disk at the WN which is critical, not that in
> the SE (which, for most large sites, will be distributing load over
> (24 to 36 disks) * ( 10 to 30 server) ).
Ah yes, and that's I think where the guidelines of 5MB/s per
core (I guess this assumed a typical 5-10 HS06 core) for worker
nodes comes in. While the 20MB/s RW per TB applies to site
global storage IIRC.
But all this was mostly for me to understand the big picture.
As to my site, we are mostly ATLAS/LHCb oriented (with 'pheno'
filling in) and strictly "production" so I guess for that the
rule is simply "more cores", or as my local users would probably
say, "more more more more more more more cores" :-).
|