| View previous topic :: View next topic |
| Author |
Message |
Gentoopc
Guru
Joined: 25 Dec 2017
Posts: 364
|
 Posted: Wed Jan 10, 2024 12:06 am Post subject: Discussion about CUDA |
 |
|
Moderator note: Split from IO Performance loss after kernel upgrades from 5.15 to 6.x -- pietinger
as a rule, I have changes in the kernel config when building new kernel versions. if I build the kernel, I change the build options. It seems to me that there is a slight increase in performance in new versions of the kernel. but the cost of this increase negates everything. maybe the new IO scheduler in kernel versions 6.7 and higher will bring significant changes. the best thing that kernel developers could do today is to shift a number of tasks, for example, to CUDA kernels. for example, if you place IO tasks on CUDA, then thanks to the huge parallelization capability and the possibility of other algorithms in solving these tasks, thanks to CUDA, it would make a huge difference. and so the visibility of changes without significant changes themselves. |
|
| Back to top |
|
 |
Gentoopc
Guru
Joined: 25 Dec 2017
Posts: 364
|
| Posted: Wed Jan 10, 2024 12:14 am Post subject: |
|
|
I am not an expert and not a fan of nvidia. but I just want the kernel to be able to give the user the opportunity to flexibly configure and get performance from everything that can give this performance. if today all CPUs with x86 architecture cannot enable each linux kernel process to work without displacement, then why not use CUDA? A huge number of CUDA cores would allow core processes to run freely without complex scheduling algorithms. |
|
| Back to top |
|
|
logrusx
Advocate
Joined: 22 Feb 2018
Posts: 2430
|
| Posted: Wed Jan 10, 2024 5:24 am Post subject: |
|
|
| Gentoopc wrote: |
maybe the new IO scheduler in kernel versions 6.7 and higher will bring significant changes. |
That's unreasonable to be expected. Schedulers are around for quite some time and they've reached their limits. Changes in behavior yes, but changes in performance? Only regressions.
| Gentoopc wrote: | | the best thing that kernel developers could do today is to shift a number of tasks, for example, to CUDA kernels. for example, if you place IO tasks on CUDA, then thanks to the huge parallelization capability and the possibility of other algorithms in solving these tasks, thanks to CUDA, it would make a huge difference. and so the visibility of changes without significant changes themselves. |
That's even more unreasonable. GPU's are capable of doing very specific tasks very fast. They can't even execute general purpose tasks. That's why there are things like CUDA - to enable execution of those very specific tasks for purposes other than video rendering. Otherwise everything would run on the GPU and we would not need CPU's for that matter.
Best Regards,
Georgi |
|
| Back to top |
|
|
Gentoopc
Guru
Joined: 25 Dec 2017
Posts: 364
|
| Posted: Wed Jan 10, 2024 6:52 am Post subject: |
|
|
You're wrong about that. this is very reasonable. due to the huge number of CUDA cores, it would be possible to give all the processes in need not to queue, not to waste processor time, not to be frozen as it was before and run immediately. all these processes in hibernation would not be needed. I think you know what I mean. if you do not see any benefit in this, then we will not argue.
it's just not clear why you're afraid to use all the features of iron. CUDA cores won't make it worse. do not use them, it is as if you had 16 cores on a Ryzen processor, of which 8 are disabled, and you would not agree to use them. also here, there are many CUDA cores that need to be used. they may be weak, but there are many of them, they can be assigned those tasks that require parallelization. thanks to the CUDA programming approach, it would be possible to bypass structures very quickly and simultaneously. previously, this was done in cycles, running through a structure in a cycle is not very cool if it is possible to cover it at one moment. |
|
| Back to top |
|
|
logrusx
Advocate
Joined: 22 Feb 2018
Posts: 2430
|
| Posted: Wed Jan 10, 2024 7:30 am Post subject: |
|
|
| Gentoopc wrote: | You're wrong about that. this is very reasonable. due to the huge number of CUDA cores, it would be possible to give all the processes in need not to queue, not to waste processor time, not to be frozen as it was before and run immediately. all these processes in hibernation would not be needed. I think you know what I mean. if you do not see any benefit in this, then we will not argue.
it's just not clear why you're afraid to use all the features of iron. CUDA cores won't make it worse. do not use them, it is as if you had 16 cores on a Ryzen processor, of which 8 are disabled, and you would not agree to use them. also here, there are many CUDA cores that need to be used. they may be weak, but there are many of them, they can be assigned those tasks that require parallelization. thanks to the CUDA programming approach, it would be possible to bypass structures very quickly and simultaneously. previously, this was done in cycles, running through a structure in a cycle is not very cool if it is possible to cover it at one moment. |
I know you hate to read but you really have to educate yourself and it works by mostly reading. Otherwise you sound ridiculous and even more so when you insist on it.
Best Regards,
Georgi
p.s. this is off topic, so please stop it.
Last edited by logrusx on Wed Jan 10, 2024 7:49 am; edited 1 time in total |
|
| Back to top |
|
|
eccerr0r
Watchman
Joined: 01 Jul 2004
Posts: 9824
Location: almost Mile High in the USA
|
| Posted: Wed Jan 10, 2024 7:36 am Post subject: |
|
|
let everyone deduct credits from a bank account without serialization...what could possibly go wrong.
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching? |
|
| Back to top |
|
|
Gentoopc
Guru
Joined: 25 Dec 2017
Posts: 364
|
| Posted: Wed Jan 10, 2024 9:04 am Post subject: |
|
|
indeed, when you have to explain obvious useful things to people, and they deny their usefulness and you have to make the obvious obvious more and then you look really stupid. you're right. you are struggling with overheating processors that have already reached the limit in both power consumption and heat dissipation, but you have CUDA cores that are not involved, which could somehow save the situation and give new opportunities in solving complex problems, and you refuse. that's right, let CUDA be idle and you will continue to get from the CPU what they can no longer give. you can't even cool them anymore. CUDA could help in some way. you need to use all the available ones. The Linux kernel should be able to work and give good performance on any hardware and in all available ways. you need to think not only about yourself but also about other users. perhaps an increase in productivity without replacing hardware would be useful to someone. Well, let everyone stay with their opinion. |
|
| Back to top |
|
|
Hu
Administrator
Joined: 06 Mar 2007
Posts: 22657
|
| Posted: Wed Jan 10, 2024 3:48 pm Post subject: |
|
|
| My opinion is that you made this exact same argument months ago. Many people explained to you why the situation is not as simple as you make it out to be. We invited you to demonstrate that it is as simple as you say. To the best of my recollection, you never delivered such a demonstration. |
|
| Back to top |
|
|
szatox
Advocate
Joined: 27 Aug 2013
Posts: 3432
|
| Posted: Wed Jan 10, 2024 5:29 pm Post subject: |
|
|
Eh... I'll make one last attempt.
| Quote: | | for example, if you place IO tasks on CUDA, then thanks to the huge parallelization capability and the possibility of other algorithms in solving these tasks, thanks to CUDA, it would make a huge difference |
Pretty much all IO use DMA anyway. DMA operations happen on the memory bus instead of CPU. |
|
| Back to top |
|
|
logrusx
Advocate
Joined: 22 Feb 2018
Posts: 2430
|
| Posted: Wed Jan 10, 2024 6:57 pm Post subject: |
|
|
| szatox wrote: | Eh... I'll make one last attempt.
| Quote: | | for example, if you place IO tasks on CUDA, then thanks to the huge parallelization capability and the possibility of other algorithms in solving these tasks, thanks to CUDA, it would make a huge difference |
Pretty much all IO use DMA anyway. DMA operations happen on the memory bus instead of CPU. |
To add to szatox's replay, that is actually something I learned in school. DMA is a special mechanism that allows devices direct access to memory, while the CPU is in high impedance state and doesn't interfere with the transfer. This way memory transfer is offloaded from the CPU to the DMA controller. DMA stands for Direct Memory Access. Back in the days you even had to enable it manually in the kernel for ATA devices so that you get decent disk IO. Nowadays most of this happens behind the scenes.
Furthermore the GPU is not built to communicate with other hardware and it's build for operations that are executed in parallel on huge amounts of data, the same operation over and over again like matrix multiplication et.c.
It doesn't and cannot do general purpose operations, it sucks at sequential operations like it would be slower than Pentium, even 486 and most of the tasks are sequential tasks, at best executed in parallel which is not something the GPU is good at. It was designed that way and however smart you think you may be, you can't outsmart the design of something that was intentionally built that way.
So, Gentoopc, please educate yourself, you really need it an it cannot be done on a forum especially when you resist what you're being taught.
Best Regards,
Georgi |
|
| Back to top |
|
|
|
|
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
|
|
Gentoo Chat
