Testing HyperThreading in TMPGenc ***updated Mulittasking test*** A must see **added test 1/10/2004**

[TerryMathews]

Originally posted by: Duvie
But what you have to remember that 20-30 is pure bonus...I didn't pay for 2 cpus and I didn't have to pay fro much more expensive dual cpu motherboards or more expensive version of winxp pro....

I wasn't arguing with you. You must've missed the part where I said I was probably going to buy a HT P4 as my next box.

 

[dnoyeb]

having a hard time believing its necessary. You ca buy a car with a turbo, or buy a car just as fast with a bigger engine and no turbo.

But just because the turbo makes the car with the turbo faster does not necessarily mean it makes all cars faster. it has to be baked in the design.

if all players are on the same page, i think HT can be beneficial. But I dont think it gives one chip an advantage over another.

 

[Duvie]

Originally posted by: dnoyeb
having a hard time believing its necessary. You ca buy a car with a turbo, or buy a car just as fast with a bigger engine and no turbo.

But just because the turbo makes the car with the turbo faster does not necessarily mean it makes all cars faster. it has to be baked in the design.

if all players are on the same page, i think HT can be beneficial. But I dont think it gives one chip an advantage over another.

Hmmm...i am sorry but that was another stupid car analogy around these parts...I don't understand what the heck you are even talking about....

Who has the bigger engine??? Athlon64??? Most the reviews don't use HT enabled apps so it beats the athlon64's and fx's in multimedia with no help from HT enabled apps....And hey!! It is part of the design of the chip and it counts..i love how amd ppl like to say take sse2 and HT away and intel would be nothing..That is plain retarded....

to follow up your terrible car analogy how about an inline 6 twin turbo in a Nissan 300zx..that car has a much smaller engine in terms of displacement then my z-28 and the 350 yet it would eat my car up...The bottom line is how fast it is with all the parts and how well it gets the job done...I don't care how it gets there as long as it gets it done....

 

[TerryMathews]

Hyperthreading would provide benefit to any real-world system, but provides more benefit in situations where you have A) low IPC B) long pipeline.

Closest analogy is checkout lines at a supermarket, and it's not a good one. But, for the sake of argument, let's say you've got a really long snake line system at a supermarket, and two registers. One for credit, one for cash. Hyperthreading is like having a person go through the line and pull out the proper type of payment (operation) to fill an unused register (execution unit).

Athlons and Opterons both have short pipelines, high IPCs, and excellent branch predictors. Hyperthreading, in my opinion, would provide ~10% improvement in situations where the same on P4s provides ~30%.

It's worth noting that you could theoretically code a program that realized no benefit from Hyperthreading, or one that realized a 100% or more benefit. All it does is pull operations out of the pipeline to fill empty execution units.

 

[Tom]

Here is car analogy.

No HT is like having Corvette.

With HT is like having Corvette and minivan at same time, you go cruisin' while your wife is at the grocery store.

 

[rogue1979]

to follow up your terrible car analogy how about an inline 6 twin turbo in a Nissan 300zx..that car has a much smaller engine in terms of displacement then my z-28 and the 350 yet it would eat my car up...The bottom line is how fast it is with all the parts and how well it gets the job done...I don't care how it gets there as long as it gets it done....

Duvie, you better stick to hyperthreading benchmarks that you seem to be spending too much time on. The 300ZX Turbo has a V6, not an inline 6.

Just joking, but seriously, good work on all this hyperthreading benchmarks. However, you seem to be leading in the direction that all your hard work proves the P4 is a better solution than the Athlon64.

First of all with the P4's extremely long pipelines I am sure the performance advantage of hyperthreading is much more noticable than would be with the Athlon architecture. As I have stated before, most of the people around here that throw benchmarks around are heavily biased toward gaming, which the Athlon64 is undoubtedly the king. In any event, both the P4 and the Athlon are amazingly fast in most any situation, and only the most demanding users such as yourself will likely have a significant benefit from hyperthreading.

As for myself my 2500+@3200+ speeds on a 200MHz fsb is still much faster than anything I use it for. Even with all this new hardware coming out I don't have any plans to upgrade anytime soon.

 

[maumoon]

Originally posted by: chsh1ca
Interesting results. I'm hopeful we start to see multi-core on a single die.

Originally posted by: Accord99
Actually, I think Jeff is referring to the large benefit of HT for SETI. Using a single instance does 1 WU every 150min, with 2 instances, each instance processes 1 WU every 165min, or effectively 82min per WU, which is a massive gain. Most likely, the cause of this gain is due to the relatively poor coding of the SETI client, resulting in stalls that slows down the performance with 1 instance, but ironically enables opportunities to better utilize the P4's execution resources with 2 instances.

Actually, it is likely related to how Seti@home is written. I have noted this behaviour many times before, even on non-HT systems (I've even mentioned it before on a few boards). On my Athlon XP 2000+, Seti on Linux would do one WU in ~4:15. With two running, it would finish 2 in less than 6:00, meaning that for an extra 1:45 I could have two WUs completed. I noted similar behaviour on Dual PIII/933 boxes (running 4 processes only took about +50% longer). +50% seems to be the delta for non-HT processors.
It's still good to see that HT makes it +90-95%. It isn't necessarily only a benefit of HT, but more of simply how Seti is written. If you ask me for an explanation, I haven't got one. All I have is observed behaviour. HT should give that difference a great boost simply because of it's half multi-threaded capability.

EDIT:
By the way, just as a point, it's not as if HyperThreading is new anymore. It's not like "today's apps" weren't expected to already support it. After all, it was almost a year ago that Anandtech put out its article on the P4 3.06 w/HT.

For such a touted feature, I'd expect to have seen more games making use of it.

After reading your post I tested it on my AMD Athlon 1200 with 768MB of DDR memory. A single SETI workunit processes in about 5h10 to 5h30 CPU time (ignore exceptions) on my machine and reports as 99% efficient (with very little else running). Running two processes at the same time results in processing with approximately the same CPU time (5h10 to 5h30) however SETI reports them as 48% efficient and looking at the elapsed times I can see that it is taking slightly more time to process two together rather than two one after the other.

 

[Jeff7181]

Hey Duvie... I'm curious what would happen if you changed the priority of threads in Windows. For example, my SETI@Home client is set at the lowest priority, so I NEVER see a slowdown... SETI simply takes a backseat to whatever needs more CPU cycles. For example... set one of your encoding apps to the highest priority, then run another encoding app with normal priority and see if/how things change.

 

[Duvie]

OK I ran this on two of the above test...


I ran tmpgenc + DVD player and tmpgenc + DVD capture....This is with same TMPGenc settings and avi file

I set priority to high on the tmpgenc and left alone on the other apps....Then I set everything back to normal....then I set 2nd app to higher priority....

Tmpgenc + DVD capture....

HIgh = 5-6 sec slower and I dropped 1% of frames....
Normal = lowest time and no dropped frames
2nd app high = time was about 1-2 sec slower and no dropped frames.....

TMPGenc + DVD playing movie off of HDD

High = 4-5 sec slower and DVD player I saw a few jerks in the playback but better then how it acted with HT off in bios
Normal = lowest time and perfect dvd playback
2nd app high = 1-2 sec slower and perfect dvd playback


It appears in thsi rather short test that mucking with tmpgenc actually hurt performance and 2nd app slightly....Seems like HT may be prioritizing itself through the OS and doing a better job...Now ofcourse even with tinkering the performance degradation is very minor but the 2nd app in terms of dvd playback is harder to forgive....Also thing to mention is everything is still worlds btter then with HT off in the bios...

 

[txxxx]

I do wonder how much you people pay overall for your P4 rig. Infact, tell me, with all that expensive DDR memory

Also what comes to mind is, is that who on earth would play games whilst encoding? Makes no sense. I for one, wouldnt like the thought of having my game disrupted due to some disk access of another application.

HT is socially flawed whilst most humans (as myself) can only serially do one task at a time.

 

[sonoran]

>>>who on earth would play games whilst encoding

Maybe someone who got bored while waiting on the encode to finish would like to play a game in the meantime? The only reason people don't do that now is because they've been indoctrinated to believe they can't (because previously PC's haven't been up to the task).

While I agree that humans are largely serial-tasking creatures (especially when we're consumed in a video game), that does not imply that there is never a situation where we want a PC to accomplish multiple tasks at once. Why are Intel and AMD *both* looking at dual-core designs???

 

[Duvie]

Originally posted by: txxxx
I do wonder how much you people pay overall for your P4 rig. Infact, tell me, with all that expensive DDR memory

Also what comes to mind is, is that who on earth would play games whilst encoding? Makes no sense. I for one, wouldnt like the thought of having my game disrupted due to some disk access of another application.

HT is socially flawed whilst most humans (as myself) can only serially do one task at a time.

My wife may say that only men are serial taskers!!!

You never talk on the phone and drive at the same time?? You never talk on the phone and walk??? PPL multitask daily....

The fact is sonoran has it exactly rght....What I am learning makes me more inclined to game a bit as I am waiting for an encoding to get done or even now let seti run 100% full time while I do most of my current duties....I learned I can capture DV camcorder or even encode to mpeg2 on the fly why I do a majority of things without a single dropped frame....

 

[Megatomic]

Originally posted by: Duvie
My wife may say that only men are serial taskers!!!

You never talk on the phone and drive at the same time?? You never talk on the phone and walk??? PPL multitask daily....

The fact is sonoran has it exactly rght....What I am learning makes me more inclined to game a bit as I am waiting for an encoding to get done or even now let seti run 100% full time while I do most of my current duties....I learned I can capture DV camcorder or even encode to mpeg2 on the fly why I do a majority of things without a single dropped frame....

OT: People may "multitask" but most of the time it is illogical and unsafe to do so, especially in your examples. Driving and talking on the phone go together like driving and drinking alcoholic beverages.

Remember, just because you can do something doesn't mean you should do it.

 

[Duvie]

Very true in certain instanes...Hands-free kits like mine with a cell phone able to dial by word of mouth makes it more safe...

However the point being made is that just cause you drive and you push the brake does not mean your eyesight stops working or becomes impaired, maybe stuttering or blurred....Your moutht doesn't stop working as well...

While driving a person multitask quite a bit....A person is visually looking out scanning the next move while the brain tells the hands and feet to work according to desired action...Then the radio is playing and you are listening to Michael Savage!!! Your ears work and you can listen,comprehend, and remember topics discussed...Heck you may even sart having a conversstion with a passenger in the vehicle..Another separate action from the main task. You still put main priority (at least most do) on driving which sometimes slows the conversation down but you can do multiple things at once....

Maximizing time is a reason most multitask in their lives...I need to maximize my limited hours a weeknight on the PC to do as much as I can...Continue to convert my Divx titles to DVD, Work on my CADD programs, listen to the news on my tuner, back up an occasional DVD, surf the net and talk to you hosers!!!

 

[DAPUNISHER]

Edited for lack of it's topical nature.

 

[rogue1979]

I think people need to figure out that it takes more concentration to "listen" to a phone conversation than it does to talk to a person sitting next to you in a car. There probably isn't scientific data pointing this out, but if you think about it it is pretty obvious. If you are talking on the phone and someone in the room talks to you at the same time , it breaks your concentration and it is hard to listen to both. If you are talking to a person in the room and another person in the room also talks to you at the same time, it is much less confusing to deal with and most people can handle this with some success.

I can only guess that without visual support to help an audio conversation it takes more effort to concentrate on. It is a known fact that information in written form is much easier to remember than information that is received by voice or audio. Maybe this is related?

 

[UCSDHappyAsian]

thanx

 

[Rabmash]

I should bench my P4 2.8E, the 1MB cache would probably help a bit too. Glad to see positive research on HT tech though

 

[cmdridq]

This is the only HT thread I've noticed, so I thought I'd ask you guys a question. My system:
ABIT IC7-G 533/800 FSB
P4 3.2G HT 512K 800FSB
2GB DDR400 PC3200
1-80GB Maxtor SATA 8MB
1-160GB Maxtor SATA 8MB
Win 2K Pro SP4

I think I've done most of the obvious things to optimize HT, but I've got a job that takes over 30 hours to run on my system, and I'm trying to figure out if there is anything I can do to speed it up. Basically, the job analyzes an input file of over 70 million records and calculates statistics to find correlations between the records. It gets CPU bound a lot. Does anyone have any suggestions? Do you think changing to XP would make much difference? Thanks.......

 

[Accord99]

Can your job be split into two tasks, ie split the input file into two files with 35million records each. Then run two jobs at the same time and then consolidate the statistics at the end in some fashion. I would assume the CPU would be spending a lot of cycles waiting for data from the input file in your job so HT should have a noticeable benefit, if you can parallelize your job.

 

[cmdridq]

It would not be practical to split the input file and try to process it in parallel and merge the output. I'm using HT now, and most of the delay comes from waiting on the cpu and the internal cache. The disk drives are not causing the bottleneck, else I would RAID them. ???

 

[Accord99]

If you've got the time, then the easiest way to see if HT has any benefits would be to run two jobs at the same time (preferably on separate HDs to avoid disk thrashing) and see how much longer it takes. If it takes twice as long, than HT probably won't help, if it takes less than twice, then you get a general idea of HT's performance advantage and see if its worth your time to figure out a way for your job to take advantage of HT.

Your job may not be HD throughput bottlenecked, but it can still takes milliseconds to get data from the HD to the CPU, which is millions of wasted cycles. Cache misses can cost tens to hundreds of wasted cycles, and both cases are areas where HT is usually beneficial.

 

[cmdridq]

Not sure I follow you. I'm running the system now with HT enabled, and it runs faster than with it disabled. I'm wondering if there is anything else i can do to help it? One thing I was considering was adding another 160GB HDD, and setting up the 2-160's in a RAID, and leaving the OS on the 80GB HDD. I'm not sure my MOBO HDD controllers will support that, what if any other HW I would need, or if it would even be a good idea. Also, I was wondering if switching to XP Pro would make any real difference?

 

[Accord99]

Is your app multithreaded already? HT's benefits really only come into play when you have two threads that are doing significant work. I suggested running two jobs concurrently for a test because if your app isn't already multithreaded then the test would give you a reasonable estimate of the benefit of HT, if any. Also, XP is the better OS for HT processors, but some applications show no difference. Here's a German review regarding this matter:

http://babelfish.altavista.com/babelfish/urltrurl?lp=de_en&url=http://www.computerbase.de/article.php?id=229

Looking at the IC7-G specs, it should have the necessary hardware to support RAID with SATA drives. How big is your input file in megabytes, and how is it stored? Database, or just a simple flat text/binary file?

 

[Duvie]

cmdridg....YHPM...