AMD EPYC Performance Testing with systemd

https://www.percona.com/blog/2018/07/11/amd-epyc-performance-testing-ps-systemd/

by josephscott • 8 years ago

120 76 8 years ago

et2o4• 8 years ago

This is perhaps tangential to the article, but what is the advantage of running the MySQL server with more than 48 threads on a 24-core, 48 thread CPU anyway?

The fact that performance is in the EPYC cpu increasing on the older 4.13 kernel when you use 100 instead of 40 threads is surprising to me. On the Epyc (Kernel 4.15 | Ubuntu 18) and and Xeon CPU you can see it stalls or decreases from ~48 threads upwards.

Alupis• 8 years ago

I would assume it's because not every thread may be busy at the same time - some may be waiting for I/O, or simple keep-alive threads that only need to do things once in a while.

igor47• 8 years ago

running more threads than CPUs is useful whenever your workload is not 100% CPU-bound. a DB workload is likely IO-bound, so many threads will just be sittin' there waiting to be woken up a lot of the time

blattimwind3• 8 years ago

no IO is performed

Kenji• 8 years ago

Caching effects. If you have an x-thread CPU and you split your workload into y >> x parts, it might happen that suddenly a workload fits into a faster CPU cache level (e.g. L1 instead of L2) and then you see huge speedups. I've seen things speed up manifold when I had 67 threads and added one more even though it was an 8-thread CPU. Keep that in mind when you do threading.

Filligree• 8 years ago

Given it depends on the kernel version, I would imagine some kind of platform-specific tuning. Epyc is a NUMA platform, so perhaps improvements to the NUMA scheduler?

That's pure guesswork, though.

evanelias2• 8 years ago

The thread count in these graphs is typically on the client side -- it's the number of concurrent client threads (connections) coming from the benchmark. So with more client threads than CPU cores/HT threads, the benchmark can show the results of increasing the amount of CPU contention/saturation.

On the server side, MySQL defaults to using a model of 1 thread per connection, plus various additional background threads for I/O, listening for new conns, replication, signal handling, purging old row versions, etc. Most of these are configurable, but it's not as simple as "running the MySQL server with N threads". Basically, if the benchmark is using N threads on the client side, then you can assume the server actually has N + M threads, as the server-side thread count is dynamic based on the workload.

et2o4• 8 years ago

Ah, this is the answer that makes the most sense. Thank you.

viraptor3• 8 years ago

Most likely the change is from cgroups, and not systemd itself. This could be verified by booting with and without cgroup_disable=cpu.

zlynx• 8 years ago

And running systemd-cgtop might help see what's going on.

v4tk• 8 years ago

cgroup_disable=cpu and as matter of fact cgroup_disable= all cgroups improves things (adds extra 12% for bash), but not to the full extent.

dijit• 8 years ago

I'm not an expert, I'm a lowly systems admin-

But I imagine the difference is that size of systemd itself and how mysql does fork().

SystemD is 1.5MB itself on my systems where I have it, but upstart (for example) is 148KB on centos 6.

Since an AMD Epyc has roughly 64Mb of L3 Cache, a larger binary would not have to be evicted from L3 cache as often.

One of Intels generally powerful all-rounder CPUs (2687Wv4) only has 30Mb of "Smart-cache" (which is fancy speak for; not that much)

A complete guess on my part though..

plantain• 8 years ago

L1/L2/L3 Cache doesn't cache whole binaries, just hot pages from it.

lederhosen• 8 years ago

not hot pages, just cache lines

dsr_• 8 years ago

I'm upvoting you because even though you turn out to be wrong, you appear to have been sincere about it. You shouldn't be punished, even in fake internet points, for offering a hypothesis in good faith.

michaelmrose2• 8 years ago

I don't think imaginary internet points serve entirely or even mostly to punish.

It served to highlight valuable content and in some venues to weed out low value content entirely. This is more significant in large bodies of comments. Example highlighting the most useful 20 comments out of 700.

Interestingly sometimes wrong information can lead to good threads where people provide helpful corrections. In this case the thread is valuable for stirring up discussion even if the information may be wrong.

michaelmrose2• 8 years ago

An additional point it may be sometimes be more worthwhile to downvote incorrect information only if it seems unlikely to bear useful fruit and after providing enlightening information to the commentor if possible.

evanelias2• 8 years ago

how mysql does fork()

mysqld uses a single-process architecture; once running, it's not calling fork(). It creates a thread per connection, rather than forking a new process per connection like postgres.

bhouston2• 8 years ago

Fun fact: a lot of user benchmarks of Ryzen Threadripper have lower averages than standard performance. For example passmark and userbench. This is because many are running TR with just 2 dimms rather than 4 dimms. This basically reduces performance by +20%. I almost think that AMD should have just not made TR work in this case because the lower than accurate benchmarks hurt then.

I do not know exactly why it has this performance characteristic but I've witnessed it first hand. It is very ease to reproduce with Passmark on Windows.

NullPrefix2• 8 years ago

4 dimm requirement seems kind of asinine. Normal systems can boot and work fine with a single stick of ram.

bhouston2• 8 years ago

Well for high performance systems requiring 4 dimms isn't bad as the tr series has 4 memory controllers.

tankenmate2• 8 years ago

You don't have to run with 4 DIMMs but you'll run slower if you do; it's kinda like having a one lane road vs a 4 lane road, you can put all your traffic through one lane, but it's a whole lot slower than using four lanes.

CitizenKane• 8 years ago

It could be the overall memory bandwidth. Being able to write to various sticks at the same time could do a lot, especially in a CPU with this many cores

kinghajj• 8 years ago

It's definitely a memory bandwidth issue, since ThreadRipper 1 has two dies, each of which have direct access to two DIMMs. So if you only have two DIMMs total, one die has to go through InfinityFabric to the other for all memory accesses.

spamizbad2• 8 years ago

Who coughs up boku bucks for a Threadripper and an X399 board and then cheaps out on RAM?

dman• 8 years ago

It does feel weird paying more for RAM than you do for your CPU.

elorant• 8 years ago

Well, let's see. A TR4 mobo can host 8 dimms. Say you want to max out in RAM that's a minimum of $1100. Add at least 10% if you go for ECC modules. Eventually you end up paying for RAM more than what you pay for the CPU. That's why a lot of people are conservative with RAM when they build a new rig. Somehow it doesn't feel right to pay that much money for memory. So you buy the bare minimum and then wait for prices to normalize, which seems like taking forever in this freaking industry.

derefr• 8 years ago

s/boku/beaucoup/

Interesting eggcorn; where'd you learn it that way?

bpchaps• 8 years ago

You might be surprised. A lot of us early adopters had a lot of issues with getting RAM to run at its advertised clock speed. Buying expensive RAM didn't really have benefits over cheap RAM.

It took something like 4-6 months for a BIOS update to finally bring my TR's RAM to advertised clock speeds (and CPU RAID....). Though, it still had stability problems and overclocking still isn't worth the hassle. Getting the machine to a stable state was very reminiscent of 2008.

Allegedly the second gen threadrippers are going to be much more stable. heh.

darklajid• 8 years ago

Is this a Threadripper or general Ryzen thing? Got this Asus RoG GL702ZC here, didn't upgrade and it comes like that out of the box (16G, just one socket filled) afaik.

blattimwind3• 8 years ago

Desktop platforms (Intel 11xx, AMn) generally have two memory channels. So if you're only running with one module, you're probably leaving performance on the table, regardless of platform.

Chilirose2• 8 years ago

I cant get my 3200 mhz Ram run with my x399 Designare it runs only at 2133 mhz AMD has a big problem there my Z170 Board runs all my ram on 3200 xmp without any trouble Both kits I have are certified for the X399 board but non is working above 2133mhz Very dissapointing

gsich2• 8 years ago

systemd not SystemD

snorrah2• 8 years ago

Irrelevant

gsich2• 8 years ago

Kinda. It does make me feel like the author doesn't have enough knowledge about it.

citilife• 8 years ago

I ran the same benchmark on my Intel box

What was the Intel box? I'm also wondering the EPYC system was configured properly.

aoeusnth1• 8 years ago

It was listed in the first few paragraphs, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

rossmohax• 8 years ago

I am pretty sure it has something to do with kernel.sched_autogroup_enabled = 1 which places process from different sessions into different scheduling groups.Bash terminal is session leader and all processes are part of it, unless they explicitly break away with setsid(2) call

nisa• 8 years ago

Found some more details here: https://www.postgresql.org/message-id/50E4AAB1.9040902@optio...

dragontamer• 8 years ago

Its the infinity fabric. AMD EPYC doesn't have 64MB of L3 cache. It has 8x8MB of L3 cache.

* If CCX#0 has a cacheline in the E "Exclusive Owner" state, then CCX1 through CCX7 all invalidate their L3 cache. There can only be one owner at a time, because the x86 architecture demands a cohesive memory.

* All 8-caches can hold a copy of the data. In the case of code: this means your code is replicated 8x and uses 8x more space than you think it does. Code is mostly read-only. With that being said, it is shared extremely efficiently, as all 8x L3 caches can work independently. (1MB of shared code on all 8x CCXes will use up 8x1MB of L3).

* Finally: if CCX#0 has data in its L3 cache, then CCX#6 has to do the following to read it. #1: CCX#6 talks to the RAM controller, which notices that CCX#0 is the owner. The RAM controller then has to tell CCX#0 to share its more recent data (because CCX#0 may have modified the data) to CCX#6. This means that L3-to-L3 communication has higher latency than L3-to-RAM communication!

-------------

In the case of a multithreaded database, this means that a single multithreaded database will not scale very well beyond a CCX (12-threads in this case). L3-to-L3 communications over infinity fabric is way slower, because of the cache coherence protocols that multithreaded programmers rely upon to keep the data consistent.

But if you run 8x different benchmarks on the 8x different CCXes... each of which were 12-thread each, it would scale very well.

-------------

Overall, the problem seems to scale linearly-or-better up to 16ish threads. (8x is 6762.35, 16x is 13063.39).

Scaling beyond 6-threads is technically off a CCX (3-cores per CCX on the 7401), but remains on the same die. There's internal infinity fabric noise, but otherwise the two L3 caches inside a singular die seem to communicate effectively. Possibly, the L3->memory controller->L3 message is very short and quick, as its all internal.

The next critical number for the 7401 is 12-threads, which is off of a die (3+3 cores per die). This forces "external" infinity fabric messages to start going back and forth.

Going from 12-threads (10012.18) to 24-threads (16886.24) is all the proof I need. You just crossed the die barrier, and can visibly see the slowdown in scaling.

--------------

With that being said, the system looks like it scales (sub-linearly, but its still technically better) all the way up to 48 threads. Going beyond that, Linux probably struggles and begins to shift the threads around the CCXes. I dunno how Linux's scheduler works, but there are 2x CCXes per NUMA node. So even if Linux kept the threads on the same NUMA node, they'd have to have this L3 communication across infinity fabric if Linux inappropriately shifted threads between say... Thread#0 and Thread#20 on NUMA #0.

That kind of shift would rely upon a big L3-to-L3 bulk data transfer across the two different CCXes (although, on the same die). I'd guess that something like this is going on, but its completely speculation at this point.

Twirrim• 8 years ago

That doesn't really answer the question: What's with the systemd performance being slow?

AstralStorm• 8 years ago

Likely bad synchronisation primitive is used, such as sleep instead of optimistic spinning or even locking. Probably not atomic variables where they can be used instead of mutexes.

Alternatively croup confinement causes threads to be migrated at random and some flat overhead.

jonathonf2• 8 years ago

I, too, read the article incorrectly the first time round. What it's saying is that for an EPYC CPU, performance under systemd is _better_. For a Xeon CPU, performance under systemd is _worse_.

So, the EPYC's architectural differences won't explain why an EPYC CPU runs slower under systemd, because it's the opposite.

newnewpdro2• 8 years ago

Implying that systemd is somehow the root-cause for this performance disparity strikes me as ridiculous.

I've noticed a pattern over the years with anyone spelling systemd as SystemD: They tend to not really know what the hell they're talking about with regards to systemd, while possessing significant bias against the project, actively searching for reasons to disparage it.

snorrah2• 8 years ago

Feel free to illuminate us with what the possible performance discrepancy is instead caused by ?

viraptor3• 8 years ago

Could be phrased better, but they're not wrong: https://news.ycombinator.com/item?id=17518518

newnewpdro2• 8 years ago

A kernel bug, that systemd exposes by attempting to utilize more of the kernel's newer features.

AMD EPYC Performance Testing with systemd

Related Stories