SMP with TMPGEnc, multiprocessor help!

[studtrooper]

Hello,
I just upgraded to a nice 2x2.4ghz (soon to be massivly overclocked) Xeon based system (check my new specs) and had a question about how TMPGEnc uses multi processors. Now, I know TMPGEnc is SMP aware, but I am getting weird results when converting. Instead of using 100% of processing power like it did with my Athlon XP system, it uses around 25-35% with hyperthreading enabled or around 55-65% without hyperthreading enabled. Suprisingly, the encoding time is the same with or without HT enabled. I don't understand, however, why TMPGEnc won't utilize both of my processors to their fullest extent? I set the Enviornmental Settings correctly to tell TMPGEnc it has complete control, but it still doesn't look like it is working right. Anyone else use TMPGEnc with a SMP system? Some answers or insights would be of great help. Hmm... maybe I should have gone with the Athlon 64 after all

[gastorgrab]

Operating system?

On operating systems like Win2k, and WinXP pro, is a different kernal loaded at the time of install depending on how many CPU's you have?

(this is a question, not a statement)

[Capmaster]

Hello,
I just upgraded to a nice 2x2.4ghz (soon to be massivly overclocked) Xeon based system (check my new specs) and had a question about how TMPGEnc uses multi processors. Now, I know TMPGEnc is SMP aware, but I am getting weird results when converting. Instead of using 100% of processing power like it did with my Athlon XP system, it uses around 25-35% with hyperthreading enabled or around 55-75% without hyperthreading enabled. Suprisingly, the encoding time is the same with or without HT enabled. I don't understand, however, why TMPGEnc won't utilize both of my processors to their fullest extent? I set the Enviornmental Settings correctly to tell TMPGEnc it has complete control, but it still doesn't look like it is working right. Anyone else use TMPGEnc with a SMP system? Some answers or insights would be of great help. Hmm... maybe I should have gone with the Athlon 64 after all

Sometimes where you set the priority in Windows affects how much CPU time, worst-case, that will be allowed. You may need to tweak it.

[studtrooper]

Operating system?

On operating systems like Win2k, and WinXP pro, is a different kernal loaded at the time of install depending on how many CPU's you have?

(this is a question, not a statement)

Using Windows XP Pro with a ASUS PC-DL Deluxe.

[BJ_M]

well first - dont use hyperthreading with dual xeons ... no point really except in a few very rare cases ..

second -- tmpgenc will only use 2 cpus anyway (except older 1.2 version would use 4) ...

third -- you have to set it up in tmpgenc to use all cpus and such -- and have the memory to support it ...

check off everything in the multi thread setting and increase prefetch to 128 meg - or higher ...

also -- xeons dont ussually overclock well ... though i am not a big fan of overclocking for the most part ...

[BJ_M]

you have to use the beta BIOS 1004.012 to overclock that board -- and often have to do some mods to the board .. if you have before rev 1.05 board -- its not a good idea ..

[studtrooper]

Heh heh, believe me, I have done more than enough homework on the PC-DL. Thanks for the imput, I am going to implement this immediately

[studtrooper]

Well, I followed those tips and TMPGEnc still seems content with using approx. 60% of both processors simultaneously. Is this just me, or does this happen to others too?

[rallynavvie]

You aren't using v3 are you? I'm using 2.5 and it works fine with both my SMP rigs. Everything is checked under the CPU tab as it should be. Is it detecting two logical CPUs to the right in that tab? May also be your OCing. Have you verified that OC with Prime95 yet?

Careful overclocking that Asus rig. You may need better RAM for that thing, I think the PC-DL OCers were using BH-5 and BH-6 chips for the best results. Are those 2.4s M0s? People are getting some crazy OCs with those on the DH800. Mmm, dual 4-phase VRM *drool*

[studtrooper]

Its not that you need "good ram" per say if you are not intending to go over 166fsb, its just that the PC-DL has a bad habit of undervolting the memory and CPU/S, plus there are no BIOS voltage adjustments available. I think my CPUs are rated for 1.5v and they get only around 1.46v. If I wanted a higher overclock on either my mem or CPUs, I'd need to do a little modding to my mobo to compensate for the voltage droop.

As for your first question, yes I am using TMPGEnc 2.5 and it detects 2 logical processors. I refuse to use 3 because they force invasive Online checking, absolutely unacceptable. Its not that I'm having a serious problem with TMPGEnc, my encodes come out great, its just that it doesn't seem to want to utilize both processors like it does a single processor. This was pretty much the main reason I upgraded to a Dual CPU setup for faster TMPGEnc encodes... I think my Athlon XP 2500+ Barton @ 2300mhz was a little bit faster than my current setup (see my "Computer Details")...

Using two 2.4ghz Xeon Engineering Sample CPUs. If I can up the voltage here I think I can hit 3.5ghz easy...

[BJ_M]

it also depends on the source -- if tmpgenc has to do more decoding or if the source feed is slow -- will slow down your numbers ..

some filters seem also to use only one cpu

[studtrooper]

Crap. Well, I just loaded up the latest DVD Shrink and to my surprise, it loves SMP. 100% full load compressing dvd files. Good Lord it is fast, probably 2.5x faster than my previous system...

[rallynavvie]

*sigh* The PC-DL had so much potential. What is its Nocona-supporting successor, the NCCH-DL? That one may be fun

All I need is a BSEL mod to get higher than 200 FSB on D1s though

However there are some very interesting quirks about the ES chips. With those, good RAM, and some hard mods to that board you should be getting up to the 266 FSB zone some of the DH800 OCers are getting to.

Sweet, another SMP power-user about

[studtrooper]

LOL, I tried the moosemod the other day but all I got was a black screen and a badly synthized voice saying "CPU Detect Failure". Might need more voltage for both the processors and mem to get 200fsb and beyond. Ahh... the joys of being a nerd

Bad thing about the NCCH-DL is that it is $100 more than the PC-DL plus it only has 2 regular PCI slots (3 useless PCI-X slots :/ ). The mods are worth it to me for saving $100 and having 5 PCI slots

[SaSi]

Going back to the original question, I have noticed that with dual CPU systems or HT enabled on a single P4, CPU utilization figures are fake.

I noticed that first with VirtualDUB on my single 2.8HT P4.

With HT enabled, I got no more than 50% utilization and approx. 32fps

Same file same settings with HT disabled gave me 100% utilization and 32 fps again.

It seems to me that when a single program is utilizing a single thread of the CPU (as virtualDUB does), windows only show 50% (half of the CPU threads are utilized) but the whole CPU horsepower is put to work.

You should measure whether Tmpgenc utilizes both CPUs by measuring the encode time of a short clip using known settings.

[BJ_M]

HT is kinda fake anyway -- thats the point .. it doesnt really help most apps at all and it is said that it should never be used on dual systems except in a few rare cases ....

[rallynavvie]

Yeah, I never saw the need for it in the first place after having the dual MP system. However there are rumors of some Adobe apps utilizing SMP HT for better performance than just SMP or HT alone. I haven't seen anything here to back that up though...

What's wrong with PCI-X slots? They work just the same as 32-bit PCI. What you're paying more for is the Nocona support, the PCI-X bridge, and the better VRM. Still cheaper than the DH800 for whatever reason

[studtrooper]

PCI-X arn't backwards compatable with PCI, are they? I forgot about 64-bit Nocona support and to mention the better VRM, I wrote that reply really late last night

[rallynavvie]

Unfortunately yes, PCI-X is compatable with 32-bit and up PCI tech. Problem is that on most boards the PCI-X bus is seperate from the PCI bus so you can get the faster speeds from the PCI-X slots. But if you put anything in them that isn't PCI-X it slows down that bus to whatever the lowest tech PCI card is on that bus. I saved my PCI-X slots for SCSI RAID adapters, or rather just one. I haven't decided what to put in the other one, maybe a FireWire800 controller.

[studtrooper]

Well, crap. If I knew that about the PCI-X I probably would have got the NCCH-DL over the PC-DL. Oh well, not too big of a loss...

[studtrooper]

After looking at the PCI-X slots on the NCCH-DL (pictured below), I would say that it looks impossible to fit regular PCI cards into a PCI-X slot. Maybe the PC-DL was the right choice after all

[rallynavvie]

Look the same as the ones I have on mine.

Look at the fingers on your PCI cards.

[BJ_M]

64 bit pci cards fit into pci-x card slots .... i think

not a lot of 64bit pci cards though -- i have a couple 64bit scsi cards and raid cards

i may be wrong

[ViRaL1]

If I'm not mistaken, PCI-X basically IS 64-bit PCI. I'm pretty sure it's all there was until PCI-E.

[studtrooper]

Well, I turned off the filtering I usually use with TMPGEnc and found a little suprise, CPU utilization went up to a 60-75% average for both processors. I guess there are filters that just don't work that well with SMP...

[BJ_M]

Well, I turned off the filtering I usually use with TMPGEnc and found a little suprise, CPU utilization went up to a 60-75% average for both processors. I guess there are filters that just don't work that well with SMP...

I said that already here https://www.videohelp.com/forum/viewtopic.php?p=1035252#1035252

[xxgriff]

software used must be coded to take advantage of optimizations. true for MMX/SSE/HT etc. while HT makes a nice advertising tool, the real world benefits are considerably less than industry PR would like us to beleive.

SMP is simular in that software must be able to make use of multi-threaded OS, and thus multiple CPU's. In MS NT kernel(XP/2k/NT) the priority given to an executable can prevent or allow more resources to be dedicated. but i have found that many people who go SMP, are suprised to find that having 2 CPU's not not translate into half the time or twice the speed.

Have you benched your SMP rig? just curious as to performance of your xeons for encoding.

[rallynavvie]

Mine seems fast

[studtrooper]

To be honest, while it has been a cool learning experience, if I knew what I know now, I'd probably have gone with a high-end Athlon 64 system. Seeing that TMPGEnc refuses to use 100% of my processor power like they claimed was just really disapointing. Oh well, I'll wait until dual-core Athlon 64 processors and AMD PCI-E/SLI enabled mobos are out for my next upgrade

[Capmaster]

software used must be coded to take advantage of optimizations. true for MMX/SSE/HT etc. while HT makes a nice advertising tool, the real world benefits are considerably less than industry PR would like us to beleive.

SMP is simular in that software must be able to make use of multi-threaded OS, and thus multiple CPU's. In MS NT kernel(XP/2k/NT) the priority given to an executable can prevent or allow more resources to be dedicated. but i have found that many people who go SMP, are suprised to find that having 2 CPU's not not translate into half the time or twice the speed.

Have you benched your SMP rig? just curious as to performance of your xeons for encoding.

The old formula used to be that 2 CPUs would yield the square root of two performance multiplier, or 1.4 times the power of one processor due to housekeeping, bus routing, cache sharing, etc. But I suspect that is no longer a valid number. It's probably closer to 1.8, if I had to guess. That's due to the fact that increases in hardware speed have outpaced demands placed on it by software and OS'es.