I’m not sure if anybody has seen this before, but I was testing my 100 Gbps Infiniband network in Windows 10 21H2 under WSL2 and Cygwin and the speeds between each are VASTLY different between my AMD Ryzen 9 5950X system and my Intel Core i9-12900K system.
Both have 128 GB of RAM, both have a Mellanox ConnectX-4 dual port VPI 100 Gbps Infiniband card in it (MCX456A-ECAT), a direct attached cable going to the Mellanox MSB7890 externally managed 36-port 100 Gbps Infiniband switch. Both systems have a HGST 1 TB SATA 6 Gbps 7200 rpm HDD for the OS and both are running Windows 10 21H2 (fresh install).
Any ideas as to why the performance difference is so significant between WSL2 and Cygwin?
Intel was touting better storage performance for 12th gen + general Intel/Windows optimizations in general
Yeah, it’s hard for me to compare because I don’t have another relatively recent Intel system (either 9th gen or 10th gen) to compare it with.
I’m surprised that WSL2 on the 5950X is so slow compared to just running Cygwin.
Either WSL2 is heavily optimised for Intel or AMD has some work ahead of itself despite the fact that the 5950X launched a year and three months ago.
For reference, the cluster head node, writing and reading directly (locally) on the host; I got about 2.8 GB/s write and 7.6 GB/s read.
Cluster headnode is an Intel Core i7-4930K on a Asus P9X79-E WS, with 64 GB of RAM, I think it’s a Nvidia Quadro 600, Mellanox ConnectX-4, running CentOS 7.7.1908, and that’s reading/writing to/from four Samsung 860 EVO 1 TB SATA 6 Gbps SSDs in RAID5.
Yeah, the only thing that comes to mind is the I/O optimizations Intel was discussing.
I also recall that WSL2 did a complete rework of I/O stuff coming from WSL, as well, since that was a major bottleneck.
It would make sense that they take optimized hardware “cheats” for intel systems, if available.
Wouldn’t be the first time we’ve seen Intel optimized software run slower on AMD.
Honestly, it’s a little of both, probably
Yeah…I’m not sure how much weight I’d put in towards Intel’s marketing/promo material vs. the actual, tested data as shown above.
Both systems are running DDR4-3200 RAM, so that isn’t a factor/variable in this difference either.
I would have expected that since AMD was the first out of the gate to have PCIe 4.0, as well as a number of other “firsts”, that things would’ve worked better on the AMD platform than it would have on the Intel platform, due at least in part to that.
I remember reading about how vendors that were trying to certify their PCIe 4.0 devices HAD to use the AMD platform for certification because Intel, at the time, didn’t HAVE a PCIe 4.0 platform that vendors could work with.
I am hoping that someone who is smarter than I am would be able to do some investigative work and to try and maybe help to figure out why I am seeing the results that I am seeing.
(I don’t do development work, so that’s not really an issue, but I WAS hoping that I might be able to use
pixz in WSL2 because there isn’t a native Windows binary for it, but if the storage I/O speeds are going to be this bad, then that’s not really going to be a viable option, which means that it looks like that I am going to have to compile
pixz on my system in Cygwin in order to get vastly better performance.)
Well, the I/O differences in performance is actually fairly measurable; to the point that SSD reviewers I was following switched to Intel systems, but I get what you mean.
This would actually be interesting thing for @wendell and the guys to maybe look into for a video?