On Thursday, May 24, 2018 7:50:17 AM EDT Allen Rout wrote:
> I've got a friend who's interested in having a physical server built at
> a breakneck pace, with particular requirements: absolutely ridiculous
> IOPS because storage bandwidth is bottlenecking his builds.
> He wants "someone who has done this dozen of times before, can
> absolutely nail the task, and is in it for the gig -- not someone who is
> in it to learn something – this task needs to be mother’s milk for the
> person with the right experience level".
> He talks like that in person, too. :)
> Let me know if you're interested, and have built some high performance
> boxes a few times.
Without knowing the size of the project, what kind of storage bandwidth
they've got currently, what makes them think they're IO limited, what their
target bandwidth is-- it's kind of hard to even know what sort of system is
actually being targeted.
My experience is that threads and memory outweighs IO when compiling
software-- so a threadripper box with a pair of NVMe drives in a RAID 0
configuration would be where I'd start.
For comparison, that box should be able to do a default linux kernel compile
in around 36 seconds (according to Phoronix). An Epyc CPU apparently can
reduce that to about 20 seconds.
Phoronix also compared compiling on a ramdisk vs a NVMe SSD and couldn't find
a "meaningful difference", further supporting my contention that disk IO isn't
the limiting factor.
Now, if the goal is to see how many IOPS you can get on a single user system,
that's just a question of how much money are you willing to spend? :)
John Gelnaw <[log in to unmask]>
University of Florida Health IT