What causes the changes in reported CPU frequency and Geekbench 5 Performance?
Cores
MHz
GB Single core
GB Multi core
32
2185.527
1719
15294
1
2700.357
1506
1454
2
2406.790
1510
1973
4
2195.458
1461
3621
8
2194.465
1446
6332
Note that 32 was on the bare metal and 1, 2, 4, and 8 were "also on the bare metal" but via LXC. I haven't yet run yabs on KVM. With LXC, rebooting the containers to change the number of cores took less than a blink each time.
In particular,
(1) why should the single core score be better with LXC than without?
(2) As between 32 and 1 core tests, why does the GB single core performance decrease when the processor speed increases?
(3) Does the GB single core score decrease with the "real core" count? 1 and 2 logical cores might run on one hardware core. 3 and 4 might run on a second "real core."
What's going on with the IPv4 Send Speed to Clouvider in NY?
Cores
Clouvider NY IPv4 Send Speed Mbits/sec
32
94.8
1
83.1
2
121
4
151
8
48.5
Who decided on those swap allocations?
Is there more that seems worthy of notice?
I hope I didn't make too many mistakes entering the data into the tables. My data guy is off today.
Here's my amateur attempt to answer some of your questions --
@Not_Oles said: (1) why should the single core score be better with LXC than without?
Based on your numbers posted, the inverse of that is true, as expected. Your first run on the host has the best single core score overall. Unless you're referring to the raw CPU frequency, in which case see my next answer below.
@Not_Oles said: (2) As between 32 and 1 core tests, why does the GB single core performance decrease when the processor speed increases?
I wouldn't pay too much attention to the CPU freq as that is just whatever speed the OS + CPU is reporting at the time that the stat is pulled (cat /proc/cpuinfo). The reported clock will always be somewhere between the min and max frequency (inclusive). The more cores it has, the more likely some of those cores are idling and awaiting work (therefore more chance of a lower frequency being reported). Those frequencies are likely hitting the max clock of the CPU (or burst clock, if available) when it's geekbench crunching time.
@Not_Oles said: (3) Does the GB single core score decrease with the "real core" count? 1 and 2 logical cores might run on one hardware core. 3 and 4 might run on a second "real core."
This I'm not so sure of. Ideally, the single core score should remain relatively the same no matter how many cores are allocated to the VM. For more concrete answers, I'd run more tests, as to me this looks to be within the margin of error to call all the single-core scores "somewhat equal".
@Not_Oles said: What's going on with the IPv4 Send Speed to Clouvider in NY?
I'll probably just be repeating what you already know here, but throughput of the network is bound to fluctuate depending on when exactly the tests are being run, how much traffic is going over the routes between your server and Clouvider's, how pegged Clouvider's speedtest server is at that moment in time, etc. Wouldn't surprise me if those cross-Atlantic links go from congested to free in just a few minute time window.
I find it interesting that the IPv6 speed tests to Clouvider's NYC location didn't suffer the same issues that the IPv4 test did that you noted above. Perhaps the different protocols are using separate routes between the two locations and the IPv4 route is defaulting to a more congested one (just a theory).
@Mason said:
Here's my amateur attempt to answer some of your questions --
@Not_Oles said: (1) why should the single core score be better with LXC than without?
Based on your numbers posted, the inverse of that is true, as expected. Your first run on the host has the best single core score overall. Unless you're referring to the raw CPU frequency, in which case see my next answer below.
@Not_Oles said: (2) As between 32 and 1 core tests, why does the GB single core performance decrease when the processor speed increases?
I wouldn't pay too much attention to the CPU freq as that is just whatever speed the OS + CPU is reporting at the time that the stat is pulled (cat /proc/cpuinfo). The reported clock will always be somewhere between the min and max frequency (inclusive). The more cores it has, the more likely some of those cores are idling and awaiting work (therefore more chance of a lower frequency being reported). Those frequencies are likely hitting the max clock of the CPU (or burst clock, if available) when it's geekbench crunching time.
@Not_Oles said: (3) Does the GB single core score decrease with the "real core" count? 1 and 2 logical cores might run on one hardware core. 3 and 4 might run on a second "real core."
This I'm not so sure of. Ideally, the single core score should remain relatively the same no matter how many cores are allocated to the VM. For more concrete answers, I'd run more tests, as to me this looks to be within the margin of error to call all the single-core scores "somewhat equal".
@Not_Oles said: What's going on with the IPv4 Send Speed to Clouvider in NY?
I'll probably just be repeating what you already know here, but throughput of the network is bound to fluctuate depending on when exactly the tests are being run, how much traffic is going over the routes between your server and Clouvider's, how pegged Clouvider's speedtest server is at that moment in time, etc. Wouldn't surprise me if those cross-Atlantic links go from congested to free in just a few minute time window.
I find it interesting that the IPv6 speed tests to Clouvider's NYC location didn't suffer the same issues that the IPv4 test did that you noted above. Perhaps the different protocols are using separate routes between the two locations and the IPv4 route is defaulting to a more congested one (just a theory).
@Mason said: Unless you're referring to the raw CPU frequency, in which case see my next answer below.
You are right! I did intend to ask about the CPU frequency. I failed to write the question clearly! ?
the CPU freq as that is just whatever speed the OS + CPU is reporting at the time that the stat is pulled (cat /proc/cpuinfo). The reported clock will always be somewhere between the min and max frequency (inclusive).
Those clueless like me might imagine that yabs could be used to compare two different processors as well as to compare one processor in two different sets of circumstances. When comparing two different processors, knowing the maximum frequency might be useful. Clueless me imagined that the frequency yabs reported would be the maximum.
I want to learn about how the frequency report varies, both in general, and specifically how the 5950X speed report apparently increases from 2185 MHz to 2700 MHz just because of moving from bare metal to "inside" a "bare metal LXC container." I mean, since "the container has no walls," how would merely changing the cgroups and the namespaces increase the processor frequency significantly? There must be more to the story!
within the margin of error to call all the single-core scores "somewhat equal".
Got it!
speed tests to Clouvider's NYC location
I've seen this multiple times. I should do an mtr.
swap reported by yabs as 5.0 GiB
That's right about @Ympker being involved! Whenever anything truly great turns up, we can find him there!
The container used for the tests is still around, so I looked at the Proxmox web GUI to see how much space had been assigned for swap. Here's what I saw:
@Mason You are the best! Thanks for paying such careful attention to my questions! Also, thanks yet again for all your help previously -- both to me personally and to the LES community! ?
@Not_Oles said: The container used for the tests is still around, so I looked at the Proxmox web GUI to see how much space had been assigned for swap. Here's what I saw:
I've not done much with LXC but the swap in proxmox could be presented as a swap partition - that could still be added to via a swapfile on the disk itself which could top up the swap to 5gb.
For example on my desktop I have:
$ sudo swapon
NAME TYPE SIZE USED PRIO
/dev/sda3 partition 1.9G 696.5M -2
/swapfile file 8G 0B -3
You could disable CPU frequency scaling at a BIOS level which would keep the processor at max frequency, although not max boost frequency so you'd still possible see different results. I guess for fairness you could try and take CPU speed while running the geekbench test, but that would need a second yabs process I imagine.
root@test-container:~# swapon
NAME TYPE SIZE USED PRIO
none virtual 5G 0B 0
root@test-container:~# mount | grep swap
lxcfs on /proc/swaps type fuse.lxcfs (rw,nosuid,nodev,relatime,user_id=0,group_id=0,allow_other)
root@test-container:~#
On the same node:
root@Proxmox-VE ~ # swapon
NAME TYPE SIZE USED PRIO
/dev/md0 partition 16G 0B -2
root@Proxmox-VE ~ # mount | grep swap
root@Proxmox-VE ~ # mount | grep md0
root@Proxmox-VE ~ #
@Mason@Mr_Tom Based on the above, it looks like yabs is reporting the same number that swapon reports for the size of the swap. The number reported by yabs and swapon apparently can seem different from the size configured in Proxmox's GUI LXC container creation dialog.
Not technically a VPS. KVM VM deployed on a home server. Was just curious to see how it compares.
Half of the cores assigned to VM. (PN50 Mini pc form factor zen 2 4700U / 64gig / nvme). About 500 bucks worth of gear. Unsure why AES and AMD-V is disabled. Single core sucks cause it's essentially a laptop cpu. Surprised at the weak disk though
@havoc said:
Not technically a VPS. KVM VM deployed on a home server. Was just curious to see how it compares.
Half of the cores assigned to VM. (PN50 Mini pc form factor zen 2 4700U / 64gig / nvme). About 500 bucks worth of gear. Unsure why AES and AMD-V is disabled. Single core sucks cause it's essentially a laptop cpu. Surprised at the weak disk though
Is anyone aware of anything that could be done to improve the server's performance? I hope the performance could be as good as possible. Thanks!
that point being smart was pretty much self-proclaimed. perhaps RaveX could shed some light since he has same thing running. the GB4 single core is pretty good.
also again, if there's a chance to compare kernel 5.11 Vs proxmox 5.4.....
@corbpie said:
Has 1500 been passed on GB5 single?
Out of curiosity I ran another YABS (GB5 only). Close to breaking the 4 minute mile record but not quite Roger Bannister.
Multi seems to have dropped too.
@corbpie said:
Has 1500 been passed on GB5 single?
Out of curiosity I ran another YABS (GB5 only). Close to breaking the 4 minute mile record but not quite Roger Bannister.
Multi seems to have dropped too.
Geekbench 5 Benchmark Test:
> ---------------------------------
> Test | Value
> |
> Single Core | 1483
> Multi Core | 3519
> Full Test | https://browser.geekbench.com/v5/cpu/6615273
Guys! I still don't understand why there should be such a big difference in single core performance, just under 1500 inside the container, versus just over 1700 outside, of the container. The test result difference of approximately 200 is greater than 10%. Why would using cgroups and namespaces cause a greater than 10% difference in the test performance? Thanks!
To match @vyas, I ran another yabs on the node to test the consistency of the single core score on the bare metal AX101.
The GB single core last time was 1719. This time 1710.
The other thing that was consistent on this particular test were the busy signals on the IPv4 iperf tests. Looks like Sunday evening in Mexico / early Monday morning in Europe might not be the best time for yabs iperf testing of IPv4 performance. IPv6 was wide open, though!
The above figure shows the graph for 66% load CPU utilization for Mixed load operation. 66% load is determined by op rate while running mixed load operation which stresses on 11GB data in the Cassandra cluster and 150 threads are given in the Cassandra stress command. The above graph shows the average values of CPU utilization for 10 iterations. It shows Cassandra utilizes more CPU usage on Linux Containers than Cassandra on bare metal in both 100% load and 66% load cases. Highest CPU usage on Linux Container is 88.93%. Highest CPU usage on bare metal is 71.77%.
That's a 17% difference in favor of LXC over bare metal on this test! So a lot depends on the test being used.
But I still would like to know more about why the Geekbench 5 disfavors LXC versus bare metal.
Apologies for possibly distracting the thread away from @Mason's lovely yabs. ⭐
Comments
Found this when cleaning out old txt files, don't know who, what or where:
KnownHost Entry VPS Server
Sultan Muda - Amazon Store
Clouveo LAX-P2 Gift LET Provider Poll 2020
Geekbench 5
https://browser.geekbench.com/v5/cpu/6417326
Sultan Muda - Amazon Store
Hello!
Hetzner AX101
LXC container with one processor running on AX101
Now LXC container with two processors
Four cores
8 cores
A few items that stood out to me include:
Note that 32 was on the bare metal and 1, 2, 4, and 8 were "also on the bare metal"
but via LXC. 
I haven't yet run yabs on KVM. With LXC, rebooting the containers to change the number of cores took less than a blink each time.
In particular,
(1) why should the single core score be better with LXC than without?
(2) As between 32 and 1 core tests, why does the GB single core performance decrease when the processor speed increases?
(3) Does the GB single core score decrease with the "real core" count? 1 and 2 logical cores might run on one hardware core. 3 and 4 might run on a second "real core."
Who decided on those swap allocations?
Is there more that seems worthy of notice?
I hope I didn't make too many mistakes entering the data into the tables. My data guy is off today.
Thanks everybody! Greetings from Mexico! ??????️
Oh my!
Nexus Bytes Ryzen Powered NVMe VPS | NYC|Miami|LA|London|Netherlands| Singapore|Tokyo
Storage VPS | LiteSpeed Powered Web Hosting + SSH access | Switcher Special |
Here's my amateur attempt to answer some of your questions --
Based on your numbers posted, the inverse of that is true, as expected. Your first run on the host has the best single core score overall. Unless you're referring to the raw CPU frequency, in which case see my next answer below.
I wouldn't pay too much attention to the CPU freq as that is just whatever speed the OS + CPU is reporting at the time that the stat is pulled (
cat /proc/cpuinfo). The reported clock will always be somewhere between the min and max frequency (inclusive). The more cores it has, the more likely some of those cores are idling and awaiting work (therefore more chance of a lower frequency being reported). Those frequencies are likely hitting the max clock of the CPU (or burst clock, if available) when it's geekbench crunching time.This I'm not so sure of. Ideally, the single core score should remain relatively the same no matter how many cores are allocated to the VM. For more concrete answers, I'd run more tests, as to me this looks to be within the margin of error to call all the single-core scores "somewhat equal".
I'll probably just be repeating what you already know here, but throughput of the network is bound to fluctuate depending on when exactly the tests are being run, how much traffic is going over the routes between your server and Clouvider's, how pegged Clouvider's speedtest server is at that moment in time, etc. Wouldn't surprise me if those cross-Atlantic links go from congested to free in just a few minute time window.
I find it interesting that the IPv6 speed tests to Clouvider's NYC location didn't suffer the same issues that the IPv4 test did that you noted above. Perhaps the different protocols are using separate routes between the two locations and the IPv4 route is defaulting to a more congested one (just a theory).
Likely is baked into the LXC template that you used. But, for now, I'll blame @Ympker, because why not?
Hope the above helps more than it hurts!
Cheers!
yabs daddy
It's good to hear from you, too, mate! Hope you and your loved ones are doing well in these times
Take care!
As for the swap allocation: Ofc it was my doing. I like messing with ppl's swap
Ympker's VPN LTD Comparison, Uptime.is, Ympker's GitHub.
You are right! I did intend to ask about the CPU frequency. I failed to write the question clearly! ?
Those clueless like me might imagine that yabs could be used to compare two different processors as well as to compare one processor in two different sets of circumstances. When comparing two different processors, knowing the maximum frequency might be useful. Clueless me imagined that the frequency yabs reported would be the maximum.
I want to learn about how the frequency report varies, both in general, and specifically how the 5950X speed report apparently increases from 2185 MHz to 2700 MHz just because of moving from bare metal to "inside" a "bare metal LXC container." I mean, since "the container has no walls," how would merely changing the cgroups and the namespaces increase the processor frequency significantly? There must be more to the story!
Got it!
I've seen this multiple times. I should do an mtr.
That's right about @Ympker being involved! Whenever anything truly great turns up, we can find him there!
The container used for the tests is still around, so I looked at the Proxmox web GUI to see how much space had been assigned for swap. Here's what I saw:
@Mason You are the best! Thanks for paying such careful attention to my questions! Also, thanks yet again for all your help previously -- both to me personally and to the LES community! ?
I've not done much with LXC but the swap in proxmox could be presented as a swap partition - that could still be added to via a swapfile on the disk itself which could top up the swap to 5gb.
For example on my desktop I have:
You could disable CPU frequency scaling at a BIOS level which would keep the processor at max frequency, although not max boost frequency so you'd still possible see different results. I guess for fairness you could try and take CPU speed while running the geekbench test, but that would need a second yabs process I imagine.
@Mr_Tom
Same container:
On the same node:
Different container on a different node:
@Mason @Mr_Tom Based on the above, it looks like yabs is reporting the same number that swapon reports for the size of the swap. The number reported by yabs and swapon apparently can seem different from the size configured in Proxmox's GUI LXC container creation dialog.
I didn't notice this before! ?
Ok... that I can't answer
lol
Racknerd 1GB KVM VPS LEB Special
# https://github.com/masonr/yet-another-bench-script #
Sat Feb 13 18:15:03 IST 2021
HosterLabs 2GB KVM IPV6 special
# https://github.com/masonr/yet-another-bench-script #
Sat 13 Feb 2021 06:37:43 PM IST
@hosterlabs
Hostsolution 2GB 200GB
#https://github.com/masonr/yet-another-bench-script
Sat Feb 13 18:48:52 IST 2021
Basic System Information:
---------------------------------
Processor : Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz
CPU cores : 2 @ 2397.222 MHz
AES-NI : ✔ Enabled
VM-x/AMD-V : ❌ Disabled
RAM : 1.9 GiB
Swap : 0.0 KiB
Disk : 196.8 GiB
@cociu
Inception 1GB 1TB
# https://github.com/masonr/yet-another-bench-script #
Sat 13 Feb 19:10:58 IST 2021
@AnthonySmith
vserver 1GB 5GB
# https://github.com/masonr/yet-another-bench-script #
Sat Feb 13 19:02:57 IST 2021
@vserversite
Webhorizon 512MB 7GB
https://github.com/masonr/yet-another-bench-script
Sat Feb 13 19:04:45 IST 2021
@Abdullah
Debian KVM on Hetzner Ax101
Dedicated KVM Slice 512MB BuyVM Luxembourg Debian 10
taking Avoro Winter Deal 2019 for another spin, seems like it's time to put it back in production. less steal now than before:
I bench YABS 24/7/365 unless it's a leap year.
Webhosting24 10x10x10 Special
No matter how beautiful a bench looks, it can never beat nature for me.
Have you tried turning it off and on again?Facts!!
Nexus Bytes Ryzen Powered NVMe VPS | NYC|Miami|LA|London|Netherlands| Singapore|Tokyo
Storage VPS | LiteSpeed Powered Web Hosting + SSH access | Switcher Special |
The disk IO seems much lower than our other KVMs at Singapore, canyou create a ticket so I can look, or PM regd. email ?
https://webhorizon.net
ticket #7977790 made
and closed due to super fast response by @Abdullah
replied, seems like a confusion on my part. This is an OVZ NAT vps. sorry!
https://webhorizon.net
Dubai $5/m
TerraHost i3 BF 500gb HDD replaced with 2TB ssd
I’m a simple man I see gifs, I press thanks
After a lot of back and forth with @Not_Oles (for all good reasons!)
Sat Feb 20 06:04:39 UTC 2021 Basic System Information: --------------------------------- Processor : AMD Ryzen 9 5950X 16-Core Processor CPU cores : 4 @ 2192.292 MHz AES-NI : ✔ Enabled VM-x/AMD-V : ✔ Enabled RAM : 4.0 GiB Swap : 4.5 GiB Disk : 97.9 GiB fio Disk Speed Tests (Mixed R/W 50/50): --------------------------------- Block Size | 4k (IOPS) | 64k (IOPS) ------ | --- ---- | ---- ---- Read | 945.54 MB/s (236.3k) | 2.54 GB/s (39.8k) Write | 948.04 MB/s (237.0k) | 2.56 GB/s (40.0k) Total | 1.89 GB/s (473.3k) | 5.10 GB/s (79.8k) | | Block Size | 512k (IOPS) | 1m (IOPS) ------ | --- ---- | ---- ---- Read | 955.06 MB/s (1.8k) | 3.27 GB/s (3.2k) Write | 1.00 GB/s (1.9k) | 3.49 GB/s (3.4k) Total | 1.96 GB/s (3.8k) | 6.77 GB/s (6.6k) iperf3 Network Speed Tests (IPv4): --------------------------------- Provider | Location (Link) | Send Speed | Recv Speed | | | Clouvider | London, UK (10G) | busy | busy Online.net | Paris, FR (10G) | 947 Mbits/sec | 919 Mbits/sec Clouvider | NYC, NY, US (10G) | busy | busy iperf3 Network Speed Tests (IPv6): --------------------------------- Provider | Location (Link) | Send Speed | Recv Speed | | | Clouvider | London, UK (10G) | busy | busy Online.net | Paris, FR (10G) | 936 Mbits/sec | 924 Mbits/sec Clouvider | NYC, NY, US (10G) | busy | busy Geekbench 4 Benchmark Test: --------------------------------- Test | Value | Single Core | 6516 Multi Core | 19541 Full Test | https://browser.geekbench.com/v4/cpu/16050385 Geekbench 5 Benchmark Test: --------------------------------- Test | Value | Single Core | 1497 Multi Core | 4746 Full Test | https://browser.geekbench.com/v5/cpu/6597804blog archives
Shot in a ticket to have Clouvider check out their iperf servers.
Considering reducing the number of retries so when iperf servers aren't operational the script doesn't take forever to run
yabs daddy
I bench YABS 24/7/365 unless it's a leap year.
Not technically a VPS. KVM VM deployed on a home server. Was just curious to see how it compares.
Half of the cores assigned to VM. (PN50 Mini pc form factor zen 2 4700U / 64gig / nvme). About 500 bucks worth of gear. Unsure why AES and AMD-V is disabled. Single core sucks cause it's essentially a laptop cpu. Surprised at the weak disk though
pretty good actually for GB5
I bench YABS 24/7/365 unless it's a leap year.
Yeah pretty happy with it. Low power (10W idle), 8 core & loads of mem makes it a good always on proxmox machine for gitlab, pihole etc
@vyas @cybertech @Mason @Everybody
Is anyone aware of anything that could be done to improve the server's performance? I hope the performance could be as good as possible. Thanks!
^^^ the iperfs are looking impressively stable
that point being smart was pretty much self-proclaimed. perhaps RaveX could shed some light since he has same thing running. the GB4 single core is pretty good.
also again, if there's a chance to compare kernel 5.11 Vs proxmox 5.4.....
I bench YABS 24/7/365 unless it's a leap year.
Has 1500 been passed on GB5 single?
1719
Out of curiosity I ran another YABS (GB5 only). Close to breaking the 4 minute mile record but not quite Roger Bannister.
Multi seems to have dropped too.
Geekbench 5 Benchmark Test: --------------------------------- Test | Value | Single Core | 1483 Multi Core | 3519 Full Test | https://browser.geekbench.com/v5/cpu/6615273blog archives
drServer - Intel® Atom™ C2750 (4-Core, 2.4GHz) 8GB DDR3 1x 200GB SSD 13usd/m
root@sv:~# curl -sL yabs.sh | bash -s -- -9
# ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
# Yet-Another-Bench-Script #
# v2020-12-29 #
# https://github.com/masonr/yet-another-bench-script #
# ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
https://dttech.top - Personal Site.
Abnormal disk speeds...
Guys! I still don't understand why there should be such a big difference in single core performance, just under 1500 inside the container, versus just over 1700 outside, of the container. The test result difference of approximately 200 is greater than 10%. Why would using cgroups and namespaces cause a greater than 10% difference in the test performance? Thanks!
To match @vyas, I ran another yabs on the node to test the consistency of the single core score on the bare metal AX101.
The GB single core last time was 1719. This time 1710.
The other thing that was consistent on this particular test were the busy signals on the IPv4 iperf tests.
Looks like Sunday evening in Mexico / early Monday morning in Europe might not be the best time for yabs iperf testing of IPv4 performance. IPv6 was wide open, though! 
Wow! In at least one test LXC actually beats bare metal! I wasn't expecting that!
Here is Figure 9 from Reventh Thiruvallur Vangeepuram's thesis, Performance Comparison of Cassandra in LXC and Bare metal:
That's a 17% difference in favor of LXC over bare metal on this test! So a lot depends on the test being used.
But I still would like to know more about why the Geekbench 5 disfavors LXC versus bare metal.
Apologies for possibly distracting the thread away from @Mason's lovely yabs. ⭐
@vyas