What's a good RAID 0 setup for CFD?

northstrider · April 1, 2024, 12:29

Hi All,

I'm running a dual cpu EPYC 9684x workstation with a 990 Samsung NVME that clocks at 5gb/s. I'm finding that during Fluent simulations, while the station flies through the iterations, it kind of slows down on the IO. My Ansys vendor recommended to setup a RAID 0 hardware storage to speed up the IO but it looks like there's a while bunch of different options out there for that and I wanted to ask you guys what are some common RAID 0 setups that you are using for your CFD setups?

Thanks!

flotus1 · April 1, 2024, 17:10

IMO it would be quite the edge case, if you could get any speedup with even more sequential read/write speed.
Maybe check with iotop first, if writes are even remotely close to theoretical peak of the drive. After running out of SLC cache, Samsungs 990 Pro SSD can only sustain writing with ~1GB/s. If no thermal throttling occurs.
And if they are, maybe reconsider if you really need to write this much data while the simulation is running.

Let's assume for a second that sequential writes are the real bottleneck here, and we have to solve that via hardware: I would look into enterprise-grade SSDs first. Ones that can sustain their advertised write speed both short and long term.
Relatively speaking, Samsungs 990 "Pro" is not well equipped to handle sustained writes. Micron 7450 MAX and Samsung PM1735 or 1743 come to mind.
And if you still want to use striping: software RAID is the way to go here.

the_phew · April 10, 2024, 11:15

Are you talking about IO during the simulation or pre/post? If it's during the simulation, it's pretty common for CFD solvers to hiccup for a few seconds/even minutes when writing results. And it would have to be pretty optimized code to saturate the write throughput of a Gen4 SSD during those stages; you can monitor the SSD IO during writes to determine if this is truly the bottleneck. Unless you are needlessly writing full results for every iteration or something, it's pretty rare for IO to end up being more than a couple percent of your overall solve time.

If you're talking about IO during pre/post, that's usually more limited by CPU throughput than SSD speed. I routinely see my CFD post-processor pegging one CPU core at 100% during disk reads while the disk itself is nowhere need its theoretical read bandwidth.

TL;DR: I doubt that SSD's throughput is actually your bottleneck for anything meaningful.

April 1, 2024, 12:29	What's a good RAID 0 setup for CFD?	#1
northstrider New Member Artem Join Date: Aug 2023 Location: USA Posts: 8 Rep Power: 3	Hi All, I'm running a dual cpu EPYC 9684x workstation with a 990 Samsung NVME that clocks at 5gb/s. I'm finding that during Fluent simulations, while the station flies through the iterations, it kind of slows down on the IO. My Ansys vendor recommended to setup a RAID 0 hardware storage to speed up the IO but it looks like there's a while bunch of different options out there for that and I wanted to ask you guys what are some common RAID 0 setups that you are using for your CFD setups? Thanks!

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Setup for buoyantBoussinesqPimpleFoam with openings	SadBoySquad	OpenFOAM	5	August 25, 2022 11:31
setup problem	SANORANJAM	FLUENT	3	May 6, 2020 11:21
ssd raid recommendations	ahmedna8	Hardware	1	February 10, 2020 11:53
RAID 0 on 2-node cluster	Astan	Hardware	12	July 16, 2019 14:59
convergence: simulating a turbine, not getting very good residuals with a good mesh	aja1345	FLUENT	0	September 4, 2015 18:26

April 1, 2024, 17:10		#2
flotus1 Super Moderator Alex Join Date: Jun 2012 Location: Germany Posts: 3,427 Rep Power: 49	IMO it would be quite the edge case, if you could get any speedup with even more sequential read/write speed. Maybe check with iotop first, if writes are even remotely close to theoretical peak of the drive. After running out of SLC cache, Samsungs 990 Pro SSD can only sustain writing with ~1GB/s. If no thermal throttling occurs. And if they are, maybe reconsider if you really need to write this much data while the simulation is running. Let's assume for a second that sequential writes are the real bottleneck here, and we have to solve that via hardware: I would look into enterprise-grade SSDs first. Ones that can sustain their advertised write speed both short and long term. Relatively speaking, Samsungs 990 "Pro" is not well equipped to handle sustained writes. Micron 7450 MAX and Samsung PM1735 or 1743 come to mind. And if you still want to use striping: software RAID is the way to go here.

April 10, 2024, 11:15		#3
the_phew Member Matt Join Date: May 2011 Posts: 44 Rep Power: 15	Are you talking about IO during the simulation or pre/post? If it's during the simulation, it's pretty common for CFD solvers to hiccup for a few seconds/even minutes when writing results. And it would have to be pretty optimized code to saturate the write throughput of a Gen4 SSD during those stages; you can monitor the SSD IO during writes to determine if this is truly the bottleneck. Unless you are needlessly writing full results for every iteration or something, it's pretty rare for IO to end up being more than a couple percent of your overall solve time. If you're talking about IO during pre/post, that's usually more limited by CPU throughput than SSD speed. I routinely see my CFD post-processor pegging one CPU core at 100% during disk reads while the disk itself is nowhere need its theoretical read bandwidth. TL;DR: I doubt that SSD's throughput is actually your bottleneck for anything meaningful.