Linux vs Windows for NV GPUs
31 Dec 2020 7:32:09 UTC
Topic 224367
(moderation:
Same hardware excluding storage (5400RPM IDE via IDE2SATA converter HDD for Linux and SATA-III SSD for Windows):
Typical Linux result (GPU is NV 1050Ti):
| h1_0451.20_O2C02Cl4In0__O2MDFS2_Spotlight_451.75Hz_2223_1 | 513389727 | 24 Dec 2020 21:22:57 UTC | 24 Dec 2020 22:55:27 UTC | Completed and validated | 2,770 | 2,753 | 1,000 | Gravitational Wave search O2 Multi-Directional GPU v2.09 () x86_64-pc-linux-gnu |
Typical Windows result:
| h1_0479.10_O2C02Cl4In0__O2MDFS2_Spotlight_479.70Hz_2987_1 | 514660092 | 31 Dec 2020 0:22:08 UTC | 31 Dec 2020 1:17:17 UTC | Completed and validated | 1,339 | 1,319 | 1,000 | Gravitational Wave search O2 Multi-Directional GPU v2.09 () windows_x86_64 |
2 fold difference, but both configs are default with single task per GPU...
Does it known or some anomaly?
Happy New Year coming @all!
Raistmer* wrote: Same
)
Different speeds for HDD and SSD make a big difference!
Totally different Tasks: ...451.75Hz_2223... and ...479.70Hz_2987... !
Different input gives different speed!
Happy New Year coming @all!
Raistmer read the All things
)
Raistmer read the All things Navi10 thread and the GPU O2MDFS2_Spotlight thread and pay attention to the conversations talking about "delta frequency" or DF.
The larger or smaller the DF determines how long it takes to crunch the O2MDF tasks.
You have to run the exact same series of tasks with the same DF in both Windows and Linux before you can make any proclamation about which may or may not be faster.
You have .55 and a .60 DF examples in your post.
I see, non-homogenious
)
I see, non-homogenious tasks.
Thanks !
Overall what's the general
)
Overall what's the general opinion of performance (gain?) using Linux vs M$.
I am testing 2 work units per gpu currently on both my 1080 and 2080 to see if its truly beneficial.
PS - I also want to make a permanent switch to Linux.
Life's short; make fun of it!
I would say it depends on the
)
I would say it depends on the project. But many projects see a 3X or 4X improvement in running Linux apps versus the Windows ones.
I would generally state there is less system overhead when running Linux compared to Windows and all processes speed up to some degree.
I would say you should run some variant of Debian as being the easiest to install and run with the best compatibility with BOINC projects.
Beginners are usually recommended either Ubuntu or Mint. Both are Debian based. There is lots of Linux beginner help on the Web. Linux is an easy 10 minute install with the USB installers and you have a fully working desktop environment to investigate and learn about.
The USB installers even have the ability to run off the USB stick in Tryout mode without even installing to fixed storage.
If you are considering keeping a Windows environment, I recommend installing Linux to its own fixed storage device. I do not recommend using any kind of dual boot environment. That is basically deprecated with cheap SSD devices available for purchase and ease of installation. Linux does not need much storage space either.
You can install it with room to spare on a 64GB disk . . . . IF you can even find and purchase that low amount of storage these days. I think the smallest disks offered by most manufacturers is a 256GB drive version.
If you want to keep your Windows installation, add a second Linux disk and switch between them as needed with the BIOS boot order selection.
johnnymc wrote:Overall what's
)
If you review the top 50 list of highest RAC hosts currently participating at Einstein, you can see what OS, GPU hardware, and CPU hardware people are using to get those results.
You might notice that there are lots of Windows boxes on the list. There would not be any if there were a 3x or 4x advantage to Linux in running the Einstein Gamma-Ray Pulsar GPU application.
You'll have to work a little harder if you want to find a fair comparison of productivity for any of the CPU applications, or for the Gravity-Wave CPU application.
But calling any of these differences OS differences is wrong. None of these applications spends a material amount of OS-level resources compared to the resources applied to the difficult numeric application at hand. Instead you are comparing the build method used to create the applications. That can be material. For example, at Einstein it appears to me that for the last few years a crucial difference in build efficiency has resulted in GPU applications that greatly favor AMD over Nvidia hardware, as indicated by the much higher relative performance here of specific cards than seen by benchmark and other means out in the non-BOINC world.
While I run Windows machines because that is my home environment, and Einstein is just a hobby add-on, I respect that for many people here Linux is a better choice. Some kind of inherent computation efficiency advantage for this kind of work is not one of the good reasons for that choice.
archae86 wrote: johnnymc
)
Archae86 is absolutely correct. There is little difference in performance between the two OS types - provided both are setup and configured correctly for the hardware on which they run. You can have pretty much the same performance with both. In fact, he often gets better results than me because he spends more time tweaking things whereas I have far too many hosts to go down that rabbit hole :-).
I have a large fleet all running Linux, and all using very basic hardware at stock settings. I kicked off building the fleet in 2006 when I purchased at auction ~200 ex-business machines, all of which had Windows XP keys and many of which worked out of the box when powered on. They had Tualatin Celeron 1300 single-core processors, 128MB or 256MB RAM and 20GB hard disks (dated 2001-2002). Even today the bulk of my fleet still runs on those same 20GB disks. I found I could overclock the CPUs to about 1500MHz and they performed quite well.
I found installing/maintaining Windows a bit of a pain so during 2007 I converted everything to Linux (PCLinuxOS to be precise). At the time there was a 'mini' version available - a basic system perfect for crunching. They provide tools to 'roll your own' ISOs - which I do. It's very convenient to be able to do a 5min fresh install and have the DE and all the crunching stuff set up and ready to go. I keep a full copy of the repository on an external USB so even updating an older ISO is still quite fast. Swapping everything over to PCLOS and setting up the tools to automate things would be the smartest decision I ever made. However it did take many years and a lot of persistence to get to the current state of nirvana for me :-). Deciding that I really should teach myself how to write bash scripts (and to do a lot of research) were crucial steps.
In 2008/2009 I started an upgrading journey which involved keeping the cases and HDDs and replacing the board/cpu/ram - a lot of E6300, Q6600 and Q8400 processors, most of which are still in service today were used. In those days it was CPU crunching only.
Around 2011/2012 when GPU apps came along, I just started adding suitable GPUs/PSUs to existing machines, as well as building some 'latest hardware' new machines. I've also added 100x340W solar panels to the roof to provide power. I get the best output per watt by not crunching any CPU tasks and by using the cores just to support the GPUs. The solar panels paid for themselves so quickly (~2.5 years) that my next move will probably be to add more panels plus a big battery to cover the overnight usage.
Unless you already have a good grounding in Linux basics, my advice would be to ask yourself how much effort are you prepared to commit to making a change that is unlikely to make crunching go any faster.
Cheers,
Gary.
Well since the OP didn't
)
Well since the OP didn't specify what works best for this project, my response was exactly "it depends on the project".
On the Universe project Linux based on Debian 20.04 LTS or the variants are 3X to 4X faster than ANY Windows OS.
Look at its Top 40 list. You need to get to position #26 before a Windows host shows up.
Compare any of the BH-spin tasks which invariably will have Linux wingman and you can see my 3X/4X proclamation is not incorrect.
If he wanted to know about
)
If he wanted to know about some other project he'd be asking the questions there.
He mentioned running GPU workunits 2x to see if there was a benefit. He was talking about Einstein. I don't run any nvidia GPUs so I made no comment about that.
Cheers,
Gary.
I have a 970 EVO Plus 500GB
)
Dual-boot has always had side effects. I do what Keith suggested and isolate each OS on its own drive. I think the lifespan and/or speed of an SSD vs Flash Drive is to be considered depending on the project.
TBH if it weren't for my 'need' to escape reality an hour or two a day playing video games, I'd have ditched M$ long ago. I know there are really good VM and emulator options available now in linux and I will learn more about them.
TL;DR My hardware should perform optimally when assisting projects.
Life's short; make fun of it!
