Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

svnweb.freebsd.org

Post reply on HN
Print view View original

51–60 of 70 posts

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#51

Post by tarruda »

krylon wrote: post #20

I bought a custom-built Rzyen-based PC around Easter, and I have experienced some issues with it; I am not sure, though, were to put the blame (CPU, motherboard/firmware, operating system (which is openSUSE Tumbleweed)). Under heavy load, the machine has performed most gracefully. However, the machine does freeze (almost) completely when left idle for a while (usually > 1 hour). When it happens, it sometimes still re…

I had a similar problem with my ryzen 1700/asrock x370 taichi: Every time I left computer idle, when I came back it was frozen. I didn't try pinging to see if it responded.

What solved the problem for me completely was blacklisting the nouveau module for nvidia(about two weeks without a single freeze). In my case it was an option because I use an AMD gpu for linux and a nvidia for passthrough, so I have no need for the nouveau module to be loaded. BTW this is where I got the hint: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1690085

What I got from this is that it is not an nouveau driver bug, but rather a hardware bug that has a greater chance of being triggered if the module is loaded. The same thread suggests that disabling ASLR is also a valid workaround.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#52

Post by dom0 »

netheril96 wrote: post #43
paulmd wrote: post #29

Earlier quoted context omitted.

AMD's not doing a recall, it works decently enough for most applications. Their response is going to be "if it crashes your application, turn SMT off". Consider they didn't even do a recall when Phenom had a showstopping TLB bug, they shipped a BIOS patch that disabled TLB entirely. And remember, Epyc is on a new stepping of the silicon, it's possible this is already fixed on it. (Threadripper is not, however)

TLB, As Translation Lookaside Buffer? Won't memory access slow down dramatically if that is disabled?

The Phenom bug was about the L2 TLB; the L1 TLB worked just fine. It still decreased performance significantly — about 5-20 % depending on workload.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#53

Post by snarfy »

old-gregg wrote: post #38
powercf wrote: post #37

I hoped this just affected Ryzen CPUs, but this Reddit post indicates that it affects Epyc also: https://www.reddit.com/r/Amd/comments/6rmq6q/epyc_7551_minin... The first post on on AMD's community forum ( https://community.amd.com/thread/215773?start=0&tstart=0 ) is almost three months old, so AMD have known about this for a long time. If it's not something that can be fixed in a UEFI update, then it's bad news for…

So he's got a segfault every couple of minutes, wow... I've been running the same test for over 4 hours now on my Ryzen 1700 (and I've had several uneventful 30-40 minute runs before). To date, I only got one "internal compiler error: Illegal instruction" but no segfaults. Whatever it is, it doesn't affect every chip the same way.

That's kind of how interrupts work though. It's a random disruption of normal control flow. Something like this could fail a different way every time you run it.

If IRETQ is failing to RET under load then this is a huge issue. How do you even fix that? Load balance interrupt processing across cores? Throttle any process in a tight loop? It's all ugly hacks. AMD needs to fix this.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#54

Post by dom0 »

gruez wrote: post #32
paulmd wrote: post #29

Earlier quoted context omitted.

AMD's not doing a recall, it works decently enough for most applications. Their response is going to be "if it crashes your application, turn SMT off". Consider they didn't even do a recall when Phenom had a showstopping TLB bug, they shipped a BIOS patch that disabled TLB entirely. And remember, Epyc is on a new stepping of the silicon, it's possible this is already fixed on it. (Threadripper is not, however)

> Their response is going to be "if it crashes your application, turn SMT off". it happens when SMT is disabled >Epyc is on a new stepping of the silicon, it's possible this is already fixed on it. (Threadripper is not, however) that's assuming they caught this bug, which i doubt is the case because it's only discovered now rather than being documented in the errata.

I dunno, there were several issues in the SMT implementation publicized earlier, it is entirely possible that the root cause is the same or related.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#55

Post by dom0 »

ansible wrote: post #46
daurnimator wrote: post #44

Earlier quoted context omitted.

Someone was chatting to me today about picking up X5670 cpus for $50-$60. Compared to my 6600K they seem quite competitive: http://ark.intel.com/compare/83352,88191,47920

You can also buy used Xeon workstations and put in a new GPU card. The main problem is that the cooling isn't especially quiet, and they have proprietary motherboards and power supplies.

The PSUs are still essentially ATX, so that's usually not a big problem.

Form factor on the other hand... well. Sometimes they really go off the reservation there.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#57

Post by dom0 »

exikyut wrote: post #34
krylon wrote: post #20

I bought a custom-built Rzyen-based PC around Easter, and I have experienced some issues with it; I am not sure, though, were to put the blame (CPU, motherboard/firmware, operating system (which is openSUSE Tumbleweed)). Under heavy load, the machine has performed most gracefully. However, the machine does freeze (almost) completely when left idle for a while (usually > 1 hour). When it happens, it sometimes still re…

> When it happens, it sometimes still responds to pings, but nothing more; if I try to ssh into it from my laptop, I do not even receive a TCP ACK. Interesting. Pinging is handled entirely by the kernel. I wonder if the ACK code path has to enter and leave userspace before the response comes out the network card? If disabling C6 (as per other comments) doesn't fix it, one possible place to start would be 1. Something…

> Interesting. Pinging is handled entirely by the kernel. I wonder if the ACK code path has to enter and leave userspace before the response comes out the network card?

Normally there is an incoming backlog, i.e. the kernel accepts (SYN/ACKs) incoming connections before accept(2) happens; accept(2) just pops a connection off the backlog.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#58

Post by onli »

paulmd wrote: post #41
subwayclub wrote: post #31

Earlier quoted context omitted.

I got used to the idea of staying away from the cutting edge in the early 2000's. The original Athlon and Thunderbird processors were groundbreaking in performance, but consumer cooling and PSU design had not caught up with their power and heat output demands - the result was uncomfortably noisy and hot at best. The high end parts really are like Ferraris in that they do take more of your time and care to deal with i…

What do you do with it? I'm not really all that big a fan of APUs, their niche is decent integrated graphics cheap, but for low-end serveresque or office-box usage I'd go with the Pentium G4560. It has ECC too if that's your thing. Rumors are that 4C8T i3s may be coming with Coffee Lake too - obviously no efficiency step up but it'll be price competitive with Ryzen (right now the i3s and locked i5s usually aren't goo…

> Of course there's no guarantees, SKL-X turned out to be shit for gaming.

Only relatively. The performance is good, it just does not beat the i7-7700K. Which is not surprise given current games reliance on fast cores. But you can absolutely game with Skylake-X, and get high FPS.

I'd also stay away from last generation APUs. They are slow, as it's basically bulldozer with a gpu on top. Zen APUs coming out early 2018 might be a different story, and maybe they fixed some bugs till then.

It really is a pity the Pentium G4560 is getting so expensive. Whether that's a production reduction or miners buying it for their boxes, in any case it is with the higher price not really competitive against the Ryzen 3 1200.

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#59

Post by bitL »

ChuckMcM wrote: post #35

Dang, I guess I'm going to wait for the B stepping of the Ryzen :-)

Original Ryzen is B1 stepping. EPYC should be already a newer one, with ThreadRipper some say it's the old B1. So you might be looking for B3/C stepping instead ;-)

Re: Ryzen CPU HyperThreads break if 100% busy and interrupted to top of Memory

#60

Post by krylon »

tarruda wrote: post #51
krylon wrote: post #20

I bought a custom-built Rzyen-based PC around Easter, and I have experienced some issues with it; I am not sure, though, were to put the blame (CPU, motherboard/firmware, operating system (which is openSUSE Tumbleweed)). Under heavy load, the machine has performed most gracefully. However, the machine does freeze (almost) completely when left idle for a while (usually > 1 hour). When it happens, it sometimes still re…

I had a similar problem with my ryzen 1700/asrock x370 taichi: Every time I left computer idle, when I came back it was frozen. I didn't try pinging to see if it responded. What solved the problem for me completely was blacklisting the nouveau module for nvidia(about two weeks without a single freeze). In my case it was an option because I use an AMD gpu for linux and a nvidia for passthrough, so I have no need for t…

Thanks for that advice!

Unfortunately, I have an nVidia GPU, and I use the nouveau driver, because apparently using nVidia's proprietary driver on Tumbleweed is a huge PITA. Maybe I will have to bite that bullet eventually, though.

Post reply on HN