AMD Ryzen, EPYC 5~6% Faster Out-Of-The-Box With Linux 5.11

[remenic]

I maintain a bunch of EPYC VPSs running Ubuntu 20.04 on kernel 5.4.0-65-generic. Would these benefit from kernel 5.11 too, considering they're virtualized?

[intelfx]

I maintain a bunch of EPYC VPSs running Ubuntu 20.04 on kernel 5.4.0-65-generic. Would these benefit from kernel 5.11 too, considering they're virtualized?

If you employ virtualization, you should not use schedutil at all, because processes inside virtual machines are not visible to the host scheduler. Use performance or aggressively tuned ondemand.

[skeevy420]

Well, it'd be nice if the box with my Ryzen in it would arrive so I could actually take it out of the box and enjoy these performance benefits. At least it cleared customs day before yesterday.

Looking at a case that only contains a motherboard, power supply, 3 hdds, 1 ssd, and a GPU is such a tease. Yeah, my ram ain't arrived either but it's at the post office according to the tracking info. Cooler is across the room. I'm getting so antsy.

[muncrief]

I compiled Linux from git on the day the CPUFreq patch was added, and my computer wouldn't boot at all. It got past grub and then simply froze, with no errors recorded in any logs, at least that I could find.

But I have to admit that since the advent of journalctl I often can't find anything, as you usually have to know what to look for and issue a command with cryptic options to find it.

Of course I can't file a bug with no logs, so I'm just crossing my fingers and hoping the final release boots.

[Laughing1]

Now add the glibc 2.33 performance on top of this...

[polarathene]

If you employ virtualization, you should not use schedutil at all, because processes inside virtual machines are not visible to the host scheduler. Use performance or aggressively tuned ondemand.

That's something I always wondered and would like to see benchmarks on phoronix with (if only I were premium ha). Is it similar for disk I/O as well if it's going to a disk image and the host will also run the data through it's own disk I/O scheduler? There's really no benefit to alternatives, just go for throughput in a VM and have the host configured for responsiveness/latency?

[torsionbar28]

If only AMD would release some Xeon E3 competitors. EPYC 5000 anyone? The current 3000 and 7000 series naming certainly leaves that possibility open. It sure would be nice to have some socketed EPYC options in the 45w, 65w, 95w range, something that could replace the Opteron 4000 series. For many applications, official ECC support is a must, so Xeon E3 is the only option in this market right now.

PS. Yes I know about Ryzen Pro having ECC support, but 1) that's OEM only, and 2) there are zero OEM's selling it with ECC enabled.

[fanbelt]

Is there any chance of these patches being backported to 5.10? I really hope so because many people (myself included) jump from LTS release to LTS release.

[intelfx]

But I have to admit that since the advent of journalctl I often can't find anything, as you usually have to know what to look for and issue a command with cryptic options to find it.

Oh yeah, journalctl -b-1 -k (previous boot, kernel logs) is so cryptic, so weird.

If you do not know what to look for, how would you look for it in plain-text logs?

[Slithery]

Thanks Michael, fantastic work as always.