Message boards :
Number crunching :
GPU testing on low powered CPU
Message board moderation
Author | Message |
---|---|
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
I recently redid my Celeron J1900 system with a more appropriately size PSU since the 550w I was using was a bit much. Plus I went for a new cube case. http://i.imgur.com/Hey7HV6.jpg http://i.imgur.com/F5mytjx.jpg http://i.imgur.com/HzEmS6B.png I decided to pickup a PCIe x1 to x16 adapter while I was at it for my GTX 750 Ti FTW and give it a go while I was at it. So far there don't seem to be any issues running the GPU in the system off of the adapter. I have started with the x41zi CUDA50 app. I used that app when the 750 was in my dual Xeon X5470 system. Mostly my goal is to see how such a small CPU effects the GPU. Since the GPU isn't a super cruncher it may turn out to be a good pairing. When I was running the 750 in my dual Xeon X5470 I beleive it may have only run normal AR Arecibo tasks at that time. I did find a comment in a post stating it was running: 2 Arecibo at once: ~25min For the past few days the 750 has been running in the Celeron and chewing on a load of GUPPI VLARS. 2 GUPPI VLAR at once: 70-75min 1 GUPPI VLAR at once: 42-45m CPU usage for the CUDA app looks to run upwards at 15-17% sometimes. I don't have a reference for this system with Arecibo tasks but I believe that 2-3 times as long for GUPPI VLARs using the CUDA app should be about as expected? I'm going to run a short batch of CPU GUPPI VLARs for a data point and then my plan is to run the NV SOG r3420 app for comparison of the CUDA. Then maybe some tuning parameters. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
to run the NV SOG r3420 app for comparison of the CUDA. Then maybe some tuning parameters. ??? r3528 is current one SETI apps news We're not gonna fight them. We're gonna transcend them. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
to run the NV SOG r3420 app for comparison of the CUDA. Then maybe some tuning parameters. I had r3420 handy from the v0.45 Beta installer. Since I was not using the GPU recently I wasn't keeping up to date on the releases. Just grabbed r3528. The CUDA app is still finishing tasks. So I had not used r3420 yet. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Brent Norman Send message Joined: 1 Dec 99 Posts: 2786 Credit: 685,657,289 RAC: 835 |
You could take a look at my AMD 4200+ 2 core for a comparison AMD 4200 it works well with a 750Ti and no CPU task. Guppie ~ 35-40 minutes 2 at a time with SoG A sample: SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc4_2bit_guppi_57449_42148_HIP78709_OFF_0008.1383.831.18.27.213.vlar_2 00:39:16 (00:29:20) 2016-11-03 10:39:25 PM your-4dacd0ea75 2016-11-03 10:38:41 PM 1 CPU + 0.5 NVIDIA GPUs 74.7 Reported: OK + 159.43 MB 129.25 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc4_2bit_guppi_57449_41812_HIP78709_0007.2204.416.17.26.254.vlar_2 00:40:14 (00:30:58) 2016-11-03 10:25:24 PM your-4dacd0ea75 2016-11-03 10:21:24 PM 1 CPU + 0.5 NVIDIA GPUs 77.0 Reported: OK 160.59 MB 130.37 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc4_2bit_guppi_57449_42148_HIP78709_OFF_0008.1383.831.18.27.204.vlar_2 00:39:22 (00:29:59) 2016-11-03 10:12:23 PM your-4dacd0ea75 2016-11-03 10:00:22 PM 1 CPU + 0.5 NVIDIA GPUs 76.2 Reported: OK 159.91 MB 129.72 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc4_2bit_guppi_57449_42148_HIP78709_OFF_0008.1383.831.18.27.92.vlar_2 00:39:19 (00:29:46) 2016-11-03 9:45:22 PM your-4dacd0ea75 2016-11-03 9:41:22 PM 1 CPU + 0.5 NVIDIA GPUs 75.7 Reported: OK 159.40 MB 129.22 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc6_2bit_guppi_57398_MESSIER031_0018.8434.831.23.46.172_2 00:35:36 (00:27:08) 2016-11-03 9:22:22 PM your-4dacd0ea75 2016-11-03 9:19:59 PM 1 CPU + 0.5 NVIDIA GPUs 76.2 Reported: OK + 152.21 MB 122.08 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc6_2bit_guppi_57398_MESSIER031_0017.27820.831.23.46.109_2 00:36:05 (00:26:25) 2016-11-03 9:03:22 PM your-4dacd0ea75 2016-11-03 9:01:13 PM 1 CPU + 0.5 NVIDIA GPUs 73.2 Reported: OK + 153.67 MB 123.65 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc6_2bit_guppi_57398_MESSIER031_0019.20653.831.24.47.149_2 00:29:04 (00:22:49) 2016-11-03 8:49:51 PM your-4dacd0ea75 2016-11-03 8:45:51 PM 1 CPU + 0.5 NVIDIA GPUs 78.5 Reported: OK 150.78 MB 120.64 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc3_2bit_guppi_57451_25670_HIP69732_0021.19855.416.17.26.245.vlar_1 00:36:53 (00:28:52) 2016-11-03 8:25:39 PM your-4dacd0ea75 2016-11-03 8:23:15 PM 1 CPU + 0.5 NVIDIA GPUs 78.3 Reported: OK 139.86 MB 109.62 MB SETI@home 8.19 setiathome_v8 (opencl_nvidia_SoG) blc3_2bit_guppi_57451_24929_HIP63406_0019.20423.831.17.26.131.vlar_1 00:33:23 (00:26:59) 2016-11-03 8:20:15 PM your-4dacd0ea75 2016-11-03 8:16:15 PM 1 CPU + 0.5 NVIDIA GPUs 80.8 Reported: OK 141.70 MB 111.49 MB |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
You could take a look at my AMD 4200+ 2 core for a comparison Good info. Is your GPU using PCIe x16 or is it limited to less on the MB? As I'm using a PCIe x1 adapter with mine I may expect there could be a reduction in speed with a x1 connection. After switching from CUDA to SoG I was getting mostly normal AR Arecibo tasks. Which were running in 15-18min. I believe that is similar to how single CUDA tasks ran with the 750Ti in the 3.33GHz Xeon. I suspended the Arecibo I had on hand and let the GUPPI VLARs run one at a time. The run times look to be 18m45s-19m15s. At the moment I'm configuring to run 2 SoG tasks, but I have to adjust other project settings to free up another CPU core. Normally the system runs: 1 - iGPU AP 2 - PrimeGrid 1 - Free The NV GPU has been getting the normally free core. So I'm reducing PrimeGrid to 1 tasks while I run 2 SoG. Currently I am running 2 tasks via app_config.xml setting. It may turn out that -cpu_lock and -total_GPU_instances_num N could be a better option when running 2 tasks at once. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Brent Norman Send message Joined: 1 Dec 99 Posts: 2786 Credit: 685,657,289 RAC: 835 |
It has a x16 slot for the card, I believe it is v2.0 if I remember right. I run a fairly aggressive command line, but leave it at low priority since it wants 100% of the CPU - not good if not running LP. No locking or anything, just 0.5 GPU, 1.0 CPU, and let it run with prefs set to not download CPU tasks ... that get ugly if it tries running them. The main problem is memory, the XP box only has 1GB to play with, but it runs well 24/7 and I can't even remember the last time it crashed. Maybe a reboot every month or 2 if I'm moving something. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Running 2 GUPPI VLARs with SoG seems to give a pretty solid 35m run time for me. The CPU time is > than the Run Time, but I expect that has to do with the nature of the slower CPU. I could play with -use_sleep settings to see if they make any measurable difference. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13732 Credit: 208,696,464 RAC: 304 |
I'm using a PCIe x1 adapter with mine I may expect there could be a reduction in speed with a x1 connection. Depending on the hardware & application, yes. But as to how much it will impact crunching times- I couldn't begin to guess. With CUDA50 my PCIe bus load was generally 0%, the odd 1-2% blip. Running 1 WU at a time with SoG & some aggressive settings on my GTX 1070, I get PCIe bus load sustained peaks of around 15% on a *8 connection (PCIe v3.1) (My GTX 750Ti sometimes has very short peaks of 17%). If my figuring is correct, that would work out to around 120% of a *1 connection. However sometime next year PCIe v4 should see the light of day, with double the bandwidth of PCIe v3.x Grant Darwin NT |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
I'm using a PCIe x1 adapter with mine I may expect there could be a reduction in speed with a x1 connection. Since the system I'm using only has PCIe 2.0 x1 it seems like it may be an issue. Running the CUDA50 app I was seeing around 44% load on the bus from GPUz. With the NV SoG app I don't believe I have seen it go over 5%. While staying mostly in the 2-3% range. I've also been running with an empty mb_cmdline.txt. So adding some settings may drive it up. If so I'd just have to find the break even point between settings and bus saturation slowing things down. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13732 Credit: 208,696,464 RAC: 304 |
Running 1 WU at a time with SoG & some aggressive settings on my GTX 1070, I get PCIe bus load sustained peaks of around 15% on a *8 connection (PCIe v3.1) (My GTX 750Ti sometimes has very short peaks of 17%). To add to that- when crunching Arecibo WUs, my GTX 750Ti can have peaks of up to 22%; sustained periods of 20% Bus Interface Load. Grant Darwin NT |
Stephen "Heretic" Send message Joined: 20 Sep 12 Posts: 5557 Credit: 192,787,363 RAC: 628 |
. . Hi Hal, . . May I ask why such an older version of SoG. The current version in Lunatics Beta 6 is r3557. It has been rewritten to minimise tasks coming up as inconclusives. So maybe you should consider that. As a comparison I am running it on my Pentium D 930 3.0GHz rig with 2 GTX950s and it is doing OK, previously I had been running r3528 with two GTX970s, but the upper card was running too hot so I downgraded the system to the 950s. Just a thought. . . For comparison runtimes I am running doublets, and they are taking about 24 to 26 Mins for Arecibo tasks and about 32 to 35 Mins for Guppis (Blc2s). The later versions of SoG tend to be faster on Guppis, especially over CUDA. . . Just my 2 bits :) Stephen . |
Stephen "Heretic" Send message Joined: 20 Sep 12 Posts: 5557 Credit: 192,787,363 RAC: 628 |
Running 2 GUPPI VLARs with SoG seems to give a pretty solid 35m run time for me. . . Hi again, . . I am using -use_sleep to get 4 concurrent tasks on the GPus with only 2 cores in the Pentium D. . . The command in the command line file with _SoG in the name is- -high_prec_timer -use sleep. . . The Pentium D copes quite OK with the load. Stephen . |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
If you want to ask why I was going to use that version you can. Raistmer asked me the same thing and I was up to r3528 less than an hour from my OP. Since I'm currently using a NV GPU I have a reason to keep on the current release builds. In fact I had switched from r3528 to r3557 earlier today. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Running 2 GUPPI VLARs with SoG seems to give a pretty solid 35m run time for me. It looks like I had really noticed -high_prec_timer in the read me. However I'm not sure if the system I'm using the 750TI FTW in supports that hardware function. I know several of my MBs can toggle HPET off/on. It is likely HPET is supported on the system but can't be disabled. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Stephen "Heretic" Send message Joined: 20 Sep 12 Posts: 5557 Credit: 192,787,363 RAC: 628 |
. . Good to see we are on the same page then :) Stephen . |
Stephen "Heretic" Send message Joined: 20 Sep 12 Posts: 5557 Credit: 192,787,363 RAC: 628 |
Running 2 GUPPI VLARs with SoG seems to give a pretty solid 35m run time for me. . . I suspect a zero current initialisation would restore the rig to defaults. Stephen . |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Running 2 GUPPI VLARs with SoG seems to give a pretty solid 35m run time for me. I went the easy route and set the command in the app config. Since it didn't throw any errors it must support the function despite note having a setting to toggle it off/on in the BIOS for that particular system. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Running 2 GUPPI VLARs with SoG seems to give a pretty solid 35m run time for me. I should add that after a few days using -high_prec_timer -use_sleep I lowered the CPU reservation. From: <gpu_usage>0.5</gpu_usage> <cpu_usage>1.0</cpu_usage> To: <gpu_usage>0.5</gpu_usage> <cpu_usage>0.5</cpu_usage> This let me return PrimeGrid to running 2 CPU tasks tasks instead of 1. So the system is running: 1 iGPU AP task 2 NV SoG MB tasks 2 PrimeGrid CPU tasks 2 CPU cores free. 1 for iGPU & 1 for NV GPU. I might be able to get by leaving only 1 CPU core free for both the iGPU & NV GPU on the system. I am going to have to try running 3 PrimeGrid CPU tasks to see if that starts to slow down anything. It seems like a low powered GPU is a good match to a small CPU. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
iGPU build hardly required CPU reservation. It's supporting runtime CPU consumption quite low. SETI apps news We're not gonna fight them. We're gonna transcend them. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
iGPU build hardly required CPU reservation. It's supporting runtime CPU consumption quite low. The system when running only SETI@home AP tasks, 4 CPU AP & 1 iGPU AP, had no issues. When running other projects on the CPU, 4 CPU PrimeGrid & 1 iGPU AP, the iGPU would be slower. So I reserved a CPU to keep it happy. For the 750Ti with the NV SoG app I did some tests with 2 free CPU cores. This command line caused NV driver restarts: -high_prec_timer -use_sleep -hp -cpu_lock -cpu_lock_fixed_cpu 2 This command line also caused NV driver restarts: -high_prec_timer -use_sleep -hp This command line was OK, but was found to have slightly lower performance: -high_prec_timer -use_sleep -cpu_lock -cpu_lock_fixed_cpu 2 Looks like either the system, or the version driver I'm using, didn't like -hp command. with r3557. Now I have configured the system for 3 CPU PrimeGrid, 1 iGPU AP, & 2 NV SoG. System is currently sitting ~90% CPU load. So hopefully not much extra time is added for Starting GPU tasks. If it does maybe I'll switch to 2 PrimeGrid and 1 SETI MB CPU task for comparison. The Celeron J1900 is not great at MB CPU tasks. Which is why I don't normally run those tasks. Not sure how well it would do with GBT VLARs tho... SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.