Linux performance - POGS

[FindersKeepers]

Hi there - my test results of Linux performance output using POGS as a benchmark will be posted here. If anyone else has a machine of any description running boinc (command or GUI) on native Linux o/s (not V/M'd or in a virtualbox) please post your setup here. I am testing 24 different flavours of Linux and will add to it if others are running something different.

Initial results are positive -
Linux Manjaro 17.0 Xfce - Boinc GUI 7.6.33 - pacman package installed - excellent output; Intel i5 10% output gain over Windows 7 x64 Pro (no H/T)
Linux Mint 18.1 Cinnamon - Boinc GUI 7.6.31 - package installed - test just started Intel i5 (no H/T)
Linux Scientific (RHEL) v7.3 - Boinc command 7.7.0 - local git compiled - tested on H/T off only so far
others listed in the jpeg.

Have to say, while nothing has gone smoothly the POGS output on Linux is great so far. Impressed with the Manjaro result on the i5 box.

I will roll out stable setups to different intel boxes and compare the Linux results to Win 7, 8 and 10 Pro versions and tabulate the results in a spreadsheet. CPU only at the moment as that is much easier for me roll out. I have no plugin external cards (removed) as this is a raw CPU crunching test bench to compare boinc output on different O/S's. Desktops will be bare bones basic as GNOME will consume CPU cycles.

[edit1] Once the committee of one review process is complete, the intended eventual content of this thread with boinc performance comparison across different distros per attached.
[edit2] I now have the x5690 running manjaro tests; H/T off; 12cores 100% (much quieter)

[attachment deleted by admin]

[FindersKeepers]

Numbers are in for Mint 18.1 running cinnamon. Nowhere near the performance of Manjaro. Credsit average is a lot lower running this version. Both clean minimal O/S stock installs running off the same hardware.

Identical i5 machine
Linux Manjaro 22.2sec/cred 15,560/day
Linux Mint 28.0 sec/cred 12,340/day
- I am collecting system performance stats during this. Maybe a CPU throttling issue, maybe system processes.

[kashi]

Interesting results, quite a difference there.

Cinnamon is a Gnome fork I think. I read that Xfce was lighter than Gnome so have always used it with BOINC, when not using Dotsch UX.

According to some reviews LXDE desktop environment is lighter/faster than Xfce but I've never tried it. Tried installing Lubuntu once but it didn't work. Had also tried installing Xubuntu unsuccessfully previous to that. It seemed my Skylake didn't take kindly to any **buntu version so switched to Mint Xfce.

Those processes stealing crunching cycles can be be annoying and hard to get rid of permanently. Remember when I was first experimenting with MacOS there was some process continually gobbling up CPU resources slowing down crunching and it kept coming back after I terminated it. More recent MacOS versions I've tried don't have this problem fortunately.

When I used to actually boot into Linux long ago I preferred KDE versions as I found them easiest to work with plus there was a BOINC utility that worked with KDE. KDE had lots of bells and whistles though so it probably wasn't too light. Can't remember even thinking about one Linux version or desktop environment being lighter/faster than another back then, usability and user preference were the main considerations then.

[FindersKeepers]

Yes big difference - Manjaro is rocking it !!! I like the Xfce over anything else so far. GNOME and Cinnamon are bloated by comparison. Have warmed to Linux now given over 20 years of windoze. Amazing how much faster simple is but have included an array of different configs to test. Will load Mint Xfce to compare and try some Lxde flavours. Without a boinc GUI running it from command is not any more work once you know how - worked this out with RHEL first up with the help of Dingos compile script.

Here is a text file of a shell script I created to test the speed of each system using the exact same source files from a USB. I use 3 separate 120MB folders of various sized unique text files (converted medical texts) - this required dos2unix to be installed prior to running. Results are saved to folder - each system tested twice from a terminal window with nothing else running.

The attached I used on Mint (hence the prefix) - note it is unix text format so open it with notepad++ or similar.


[attachment deleted by admin]

[FindersKeepers]

Eventually this gets fully populated with data - a similar one to use comparing the system performance script above.
Doing this I can better imagine what it is like working with live munitions ...

Fun times. 

[attachment deleted by admin]

[FindersKeepers]

Stonking results on Manjaro with a second i5 this time a 4570 proc. The i5-4570 is a low power cpu with excellent single thread rating.
http://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i5-4570+%40+3.20GHz&id=1896

Side by side comparison of the same machine running Win10Pro
http://pogs.theskynet.org/pogs/results.php?hostid=802168&offset=0&show_names=0&state=4&appid=  <=Linux
http://pogs.theskynet.org/pogs/results.php?hostid=799722&offset=0&show_names=0&state=4&appid=  <Win10
- note Win10 was dumbed down - all background apps turned off, all microsuck data collection turned off, onedrive off, many services turned off

You can view all my test machines here - (lulz, they are not all separate machines as a few are resused and reformatted)
http://pogs.theskynet.org/pogs/hosts_user.php?userid=458685

For those interested in taking it for a spin, my procedurte for installing a basic Manjaro 17 Xfce onto HDD -
1. Use rufus to make boot USB from Manjaro ISO
2. Install onto HDD - automagic partitioning (hands free no configuring necessary)
3. After install (reboot if needed) attach internet and update before you install boinc. From terminal screen
 >$ sudo pacman-mirrors -g
 >$ sudo pacman -Syyu
 >$ sudo pacman -S boinc
- these each take a little while (I have slow internet)
4. Using 2 separate terminal screens, run boinc from the home directory in user mode
T1  >$ cd ~
T1  >$ boinc       <= this starts the boinc service, must be done first
T2  >$ cd ~
T2  >$ boincmgr  <= leave both terminals open; if you need another terminal to do anything open a 3rd
- boincmgr starts the GUI and it is business as usual. Done. Happy days.
- closing either T1 or T2 stops boinc.

I do not have graphics installed as I am running CPU only tests for now. I would add a graphics driver after updating. Manjaro is going through a release update soon I think.

[FindersKeepers]

3 batches of Linux comparisons confirm the move to Linux is a good on; comparing no GPU, no H/T running states for all

i5-2400 Win7Pro to Linux Manjaro Xfce = +5% incr
i5-4570 Win10Pro to Linux Manjaro Xfce = +13% incr !!! Sweeet.
x5690 Win7Pro to Linux Manjaro Xfce = +17% incr  <-- !W!T!F!

These are raw no frills installs on each one with no other software running.
Windoze systems have trimmed down services and optimised setup for performance.
Linux is straight out of the box.

{edit}
Rolled out Manjaro Xfce onto the big dog today. +20% {revised} output increase over Win7Pro on a dual E5-2680, fans are idling at 100% CPU load you could almost sleep beside it. Good for 86K/day, up from 72K. Purring like a kitten. Case closed on Manjaro - can tick that box.

[FindersKeepers]

Early indications are Scientific Linux SL7.3 (rhel) GNOME is only marginally slower than Manjaro Xfce - but will be reinstalling SL7.3 with the Xfce interface to compare.

Testing next week -
SL7.3 Xfce interface
Mint Xfce interface
Ubuntu 16.0 LTS and 17.0
Debian (if I can ever get one of the mirrors to finish downloading)

[FindersKeepers]

Ubuntu numbers are ordinary. Slower than windows on either Unbuntu 16.04 LTS or mint 18 (cinnamon). Nice interface, reminds me of windoze with an apple app store for applications. Also note the controversy over Ubuntu data collection. All in all happy to give Ubuntu flavours the flick. Feature rich, looks nice but slow and intrusive.

just ... like ... windoze (only slower)

[FindersKeepers]

After a week of cpu testing different machines with different Linux flavours the results so far are clear.

Winners over windoze - RHEL and Manjaro (probably any arch distro - trying Arch next up). Arch flavours slightly quicker giving higher gains.
Losers - Ubuntu variants - Ubuntu 16.04 and Mint (cinnamon). They look pretty but are slower in my testing.

The speed gains of arch and rhel are real for either h/t on and off, and are different for different platforms. This applies to windows also so I have not seen a general rule of thumb in what I have done. You need to experiment with H/T for your own setup - easy enough to compare the completion times (run times) per unit of credit.

A word of warning - if you have heat problems with boinc windows, boinc Linux will probably make them worse. Without H/T on has raised no problems with anything I have done - with H/T on you will need an air conditioned server rack or an icebox (jk) as the cpu cooling system gets a serious workout.

[FindersKeepers]

Results today confirm 2 front running distro's for me (so far) are Manjaro 17 Xfce and RHEL/SciLin 7.3. Watch the cpu heat load if running with h/t using stock cooling. If you have a watercooler crank it up. Without h/t everything I have tried ticks over nicely on 100% loaded linux. Seasoned hands would already know this.

For the RHEL to be close it is run from terminal with zero gui (graphical interface) installed. Manjaro (arch) benefits from Xfce which is a bare bones gui and a boinc gui version to go with it. RHEL/SciLin is a more robust platform as it is used in many labs and clusters.

What I would be inclined to do is load a gui only when you need it (logging, upload/download) and switch everything off using tty (terminal) only.

Just my 20c - have only been at this for 2 weeks. Results are worth pursuing Linux for anyone else interested.

[FindersKeepers]

Night and day between Win7&Win8 pro to Manjaro. 20% plus gains to Linux with H/T off for xeon MP's. 10% for single proc boards. If you have a crunching box on either win7/8 consider giving Arch Linux a go. If I lived in a cold room I would run H/T on with Linux but it's not worth the risk of divorce ... just watch the heat load as it is a ball(bearing) tearer.

[FindersKeepers]

Thesis proved again. Landed some cheap i5-4590 boxes that came with 2 unused windows pro licenses (win8pro and win10pro). Will keep them for posterity when friends want upgrades.

These boxes have a fast single thread CPU rating. Next to an i5-4570 it is slightly quicker (as you would expect).
Win10Pro 22.5sec/cred ave = 15,360 cred/day
Linux Manjaro 19.0sec/cred ave = 18,180cred/day

That is a 16% increase in output by switching to Linux (arch series). In a cluster these things would be dandy as 4 of them (16 cores) would performance match an e5-2680 v2 with 20 cores.
#justsaying

[FindersKeepers]

I am hoping Fedora is fast - the Xfce interface on the Fed25 x64 iso has rocking good looks. Arch/Manjaro/RedHat are the fastest boinc performers on everything I run (CPU). Ubuntu/mint is slower than windoze.

Have some quick config routines to get the penguins up and running.
Did not think this was going to be as much fun as it is. Will try Rosetta (CPU) once I adopt a favourite.

If using Xfce use the xfce4-sensors plugin to monitor temps (occasionally, not continuously). Saves you having to look for a custom plugin.

edit - {sigh} daggnabbit. early signs looking like fedora the slowest of the lot. bugger - the interface looks sleek.

[FindersKeepers]

Despite having the nicest Xfce4 interface to date, Fedora boinc was a flop initially. [UPDATED]
[See note below: this was run straight from the install version from ISO] <- this was a concern until the system upgrade; meaning a part of the standard routine for maximising cpu performance is to update microcode and system firmware, rather than run straight off a distro ISO image as not all results will be consistent. Similar to windows, however when you look at the CPU MIPS/FIPS/OPS they vary wildly from install to install.]

Using 7.6.22 client on Fed25 x64 Xfce from the Fedora repo the floating point CPU benchmarks (initially) were nearly half of what Manjaro benchmarks are for the same identical machines. Boinc completion times are waaaaaaay slower than anything else I have seen on Linux so far. Will run some tests today with a system update to see if this is a microcode/kernel issue - very sad, lovely distro otherwise.



[edit1] OK, so ... following a system update {dnf update} the boinc f/pt cpu benchmark went from 2100 to 3700 (an i5 test bed) which is in line with manjaro numbers - rerunning performance today to see if this translates into better numbers. out of the box was a fizzer - the update looks like it might make a difference. Here is hoping ... waiting on some wu's to complete
[edit2] Using 3x i5 testbeds now updated; f/pt cpu numbers went up as a result; 610mb 1250 library file update containing microcode and firmware has made the difference; performance improved enormously
[edit3] As basic as they are, I haven't fully appreciated the impact of the floating and integer benchmark numbers until just now. Looking through my logs they are the difference between all the boinc performance comparisons and vary wildly across Linux and windows flavours. Not every o/s gets the same performance out of the exact same piece of hardware - which is what I have observed but never fully appreciated. This has been about getting the highest possible MIPS/FIPS/OPS and there is not one single best method. The closest I have got is Manjaro. RHEL and now Fedora tied for 2nd. The microcode and firmware updates are critical to extracting performance - I have also seen Manjaro go backwards slightly with the recent update. Out of the box set and forget the best has been Manjaro and RHEL so far.
[edit4] New completed wu's in with updated Fedora - wow, huge difference and marginally behind Manjaro. Fed25 is now a new favourite Xfce interface. Much to like.