MS said:
Who do GB multi core scores don't multiply wrt cores. For example, if single core score of an 8 core VPS is 1000, then why the multi core score is around 4000 most of the time, and not 8000. It's usually around single core score x around half of total cores of that plan.
I see this happening more in plans with more than 6 cores. It multiples correctly as long as it's a 2 core plan, but then the ratio starts going down from one towards half with higher core plans.
Is it just throttling by host, CPU generation/quality, or something else?
Any ideas, @Mason, @cybertech, @all?
I bet my ass, they are throttled or you see 6 cores but you can only use 4, so they limit you to 400%.
Lemme run another geekbench with 4.
I see this with all kinds of hosts, with their newer and not so new nodes. And, all kinds of CPUs.
A 12 core E5 v4 plan with 700 for SC and around 4400 for MC.
A 6 core Ryzen 7XXX with 1800 for SC and around 5900 for MC.
I don’t think it’s because people are capping processors or anything like that, this seems to be more of a universal problem. I think the problem comes with how threads/hyperthreading/clock speeds work.
As host nodes get loaded, the CPU clock speed will start to drop since most modern processors generally clock lower in all-core work loads. So it could be that a few cores are clocking high, but in an all-core workload where more cores are being stressed, the overall CPU clock speed of each core drops. On top of that, threads are effectively splitting the CPU cores, which could also drop performance if a node is loaded and if all of the physical cores are utilized (iirc).
I’ve done some testing on EPYC Milan and noticed an almost 50% reduction in single core results when stressing half of the CPU threads, and even CPU steal presenting in VM’s at times, despite the hypervisor itself reporting 50% CPU being free. It’s also why I think that relying on Geekbench benchmarks is kind of a mediocre way to test performance. Even if a VPS host isn’t overselling their host nodes, the score is guaranteed to drop and fluctuate immensely.
I was thinking along the same lines.
This, plus other factors mentioned by @Neoon and @cybertech are also the reasons. Capping could a reason with at least a few hosts.
I'll do a raindog308 here, and write a little blog post on this topic.
I was thinking along the same lines.
This, plus other factors mentioned by @Neoon and @cybertech are also the reasons. Capping could a reason with at least a few hosts.
I'll do a raindog308 here, and write a little blog post on this topic.
I asked BeroHost for a refugee deal, same specs, same cpu.
As soon its delivered, we gonna put to the test.
I bet if these test plans were with 2 cores, the MC score would still be around 2000. But for 4 cores it won't be 4000, instead we might see something around 3200, and for 8 cores as you can see, it's 4000.
@Neoon said:
BeroHost refugee deal 8.99€
I guess @Advin was right or BeroHost is also capping at 4.
I have a spare EPYC 7763/7B13 node that is not in production yet, I will do some testing on there and report the results maybe. Will try various different loads and see how it impacts performance with EPYC Milan. I did some testing previously but I didn't document it, just observations since I didn't think anyone was really that interested. Maybe I can make a thread about it with some more data on how various CPU loads impacts GB6/GB5 in VM's, despite not being oversold.
MS , if there's any specific results/data you want to see (i.e. when node is loaded 75%, 50%, etc), just let me know and I can test it for you on a Proxmox install
MS said:
Who do GB multi core scores don't multiply wrt cores. For example, if single core score of an 8 core VPS is 1000, then why the multi core score is around 4000 most of the time, and not 8000. It's usually around single core score x around half of total cores of that plan.
I see this happening more in plans with more than 6 cores. It multiples correctly as long as it's a 2 core plan, but then the ratio starts going down from one towards half with higher core plans.
Is it just throttling by host, CPU generation/quality, or something else?
Any ideas, @Mason, @cybertech, @all?
I bet my ass, they are throttled or you see 6 cores but you can only use 4, so they limit you to 400%.
Lemme run another geekbench with 4.
I see this with all kinds of hosts, with their newer and not so new nodes. And, all kinds of CPUs.
A 12 core E5 v4 plan with 700 for SC and around 4400 for MC.
A 6 core Ryzen 7XXX with 1800 for SC and around 5900 for MC.
I don’t think it’s because people are capping processors or anything like that, this seems to be more of a universal problem. I think the problem comes with how threads/hyperthreading/clock speeds work.
As host nodes get loaded, the CPU clock speed will start to drop since most modern processors generally clock lower in all-core work loads. So it could be that a few cores are clocking high, but in an all-core workload where more cores are being stressed, the overall CPU clock speed of each core drops. On top of that, threads are effectively splitting the CPU cores, which could also drop performance if a node is loaded and if all of the physical cores are utilized (iirc).
I’ve done some testing on EPYC Milan and noticed an almost 50% reduction in single core results when stressing half of the CPU threads, and even CPU steal presenting in VM’s at times, despite the hypervisor itself reporting 50% CPU being free. It’s also why I think that relying on Geekbench benchmarks is kind of a mediocre way to test performance. Even if a VPS host isn’t overselling their host nodes, the score is guaranteed to drop and fluctuate immensely.
@Neoon said:
BeroHost refugee deal 8.99€
I guess @Advin was right or BeroHost is also capping at 4.
I have a spare EPYC 7763/7B13 node that is not in production yet, I will do some testing on there and report the results maybe. Will try various different loads and see how it impacts performance with EPYC Milan. I did some testing previously but I didn't document it, just observations since I didn't think anyone was really that interested. Maybe I can make a thread about it with some more data on how various CPU loads impacts GB6/GB5 in VM's, despite not being oversold.
Yeah, it's not that great. They have this VPU (Volume Performance Unit), which ranges from 10 to 120. I forgot and chose 10 vpu instead of 120. https://imgur.com/K32otis Boot volume sized at 50 GBs, configured for the Ultra High Performance level, at 120 VPUs/GB. The maximum performance for this volume is 11,250 IOPS. (50 GB x 225 IOPS/GB = 11,250.)
Comments
I don’t think it’s because people are capping processors or anything like that, this seems to be more of a universal problem. I think the problem comes with how threads/hyperthreading/clock speeds work.
As host nodes get loaded, the CPU clock speed will start to drop since most modern processors generally clock lower in all-core work loads. So it could be that a few cores are clocking high, but in an all-core workload where more cores are being stressed, the overall CPU clock speed of each core drops. On top of that, threads are effectively splitting the CPU cores, which could also drop performance if a node is loaded and if all of the physical cores are utilized (iirc).
I’ve done some testing on EPYC Milan and noticed an almost 50% reduction in single core results when stressing half of the CPU threads, and even CPU steal presenting in VM’s at times, despite the hypervisor itself reporting 50% CPU being free. It’s also why I think that relying on Geekbench benchmarks is kind of a mediocre way to test performance. Even if a VPS host isn’t overselling their host nodes, the score is guaranteed to drop and fluctuate immensely.
I am a representative of Advin Servers
I was thinking along the same lines.
This, plus other factors mentioned by @Neoon and @cybertech are also the reasons. Capping could a reason with at least a few hosts.
I'll do a raindog308 here, and write a little blog post on this topic.
I asked BeroHost for a refugee deal, same specs, same cpu.
As soon its delivered, we gonna put to the test.
Free NAT KVM | Free NAT LXC | Bobr
Hmm, interesting, I will run 2 tests on a 8 Core Xeon E5-2690 VPS, with different CPU settings, no capping, will post shortly.
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
8 core VCPU, no NUMA setting, all 8 cores assigned to 1 socket.
Second YABS, same 8 core but with NUMA enabled and 8 core spread to 2 sockets, each having 4 CPU
Prox 8.1, no capping. As there is no difference in score, I think it might be a test issue?
CPU did not pass 35% total usage on both tests, and I can confirm abusers doing 100% on 16 cores, so it is not a KVM/PROX/Setup Issue in my opinion.
Also this node is empty, so only this VPS run on the whole node.
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
Yea, it's the hyper-threading effect.
I bet if these test plans were with 2 cores, the MC score would still be around 2000. But for 4 cores it won't be 4000, instead we might see something around 3200, and for 8 cores as you can see, it's 4000.
Thanks for taking part in this little experiment.
I wonder if we'll see the same problem feature with the non-hyper-threading/SMT CPUs. Like Ryzen 3 3200G, which has 4 Cores = 4 Threads.
MS
Hmm, good point.
I will stop the server later and disable HT, redo the tests, ~2Hours
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
MS
ok, HT disabled, same node, 1 VPS on it
8 core, 2 socket, 4 cores each, numa enabled
doing the 1 socket, 8 core, no numa right now
PS: Overall usage on CPU did not go over 45% during the test.
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
well, it is kinda the same.
GB6 bug? let me run this on bare-metal node, it will take 8 minutes
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
HT enabled coming next, bare metal
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
MS
It is kinda the same, Performance on VPS is ±5-7% less then on bare metal,
GB6 score the same, with or without HT
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
BeroHost refugee deal 8.99€
I guess @Advin was right or BeroHost is also capping at 4.
Free NAT KVM | Free NAT LXC | Bobr
I have a spare EPYC 7763/7B13 node that is not in production yet, I will do some testing on there and report the results maybe. Will try various different loads and see how it impacts performance with EPYC Milan. I did some testing previously but I didn't document it, just observations since I didn't think anyone was really that interested. Maybe I can make a thread about it with some more data on how various CPU loads impacts GB6/GB5 in VM's, despite not being oversold.
I am a representative of Advin Servers
MS , if there's any specific results/data you want to see (i.e. when node is loaded 75%, 50%, etc), just let me know and I can test it for you on a Proxmox install
I am a representative of Advin Servers
dont forget crazy deals after testing.
I bench YABS 24/7/365 unless it's a leap year.
@Advin Can you compare GB5 on a VM with 8 cores assigned, SMT enabled & disabled on host node.
I don't have a spare node, but VM performance should be better with SMT disabled.
https://webhorizon.net
Yeah, for sure
I am a representative of Advin Servers
Aéza Hi-CPU (i9 14900k) | Moscow, 🇷🇺 | 8.51€/month
Oracle VM - $317.96/month
Spoiler
100 GB on raid 5/6 at 320USD/mo?
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
The $320/month covers only the compute. Oracle provides free 200GB block volume, and additional storage costs approximately $0.04/GB.
I am not amazed
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
Yeah, it's not that great. They have this VPU (Volume Performance Unit), which ranges from 10 to 120. I forgot and chose 10 vpu instead of 120.
https://imgur.com/K32otis
Boot volume sized at 50 GBs, configured for the Ultra High Performance level, at 120 VPUs/GB. The maximum performance for this volume is 11,250 IOPS. (50 GB x 225 IOPS/GB = 11,250.)
@alamin655
THX for the update, I imagined Oracle having a better price than MSFT.
Host-C - VPS Services Provider - AS211462
"If there is no struggle there is no progress"
Oracle VM - AMD EPYC™ 7742
Spoiler
HostBRR Valentine Special - EUR 4.5/m
@Labze give us some exclusive love on LES.
The Ultimate Speedtest Script | Get Instant Alerts on new LES/LET deals | Cheap VPS Deals
FREE KVM VPS - FreeVPS.org | FREE LXC VPS - MicroLXC
@sh97 that looks amazing
I bench YABS 24/7/365 unless it's a leap year.
@sh97 is your HostBRR Valentine Special - EUR 4.5/m running stable?
Niet.
There is some intermittent network issue.
The Ultimate Speedtest Script | Get Instant Alerts on new LES/LET deals | Cheap VPS Deals
FREE KVM VPS - FreeVPS.org | FREE LXC VPS - MicroLXC
EPYC 7702 / 16 x 64GB DDR4 / 4 x 3.84TB NVMe + 1TB NVMe + 2 x 512GB NVMe
225W power limit, seems to be 185W to CPU under 100% load for some reason
2 GHz all core CPU load, 308W full system power
I am a representative of Advin Servers