13 Replies Latest reply on Nov 3, 2012 5:31 AM by Harm Millaard

    Harm's Monster is FAST

    Laura Burton Gotz

      Ok, I was just looking at the results on PPBM5.com and saw that about an hour ago, Harm Millard just broke the speed record for CS6 users by 22 seconds.  Holy cow!

       

      His system came in at 123 seconds with the next fastest being 145. That is one heck of a fantastic jump.

       

      I am certain he will be happy to enlighten us as to what caused that, but as far as I can tell, he has the same GTX680 that the other 3 people ahead of me have. I have the top ranked CS6 system with the GTX670. A choice I made and that I am OK with. Some of it has to be that RAID30. I didn't realize that his drive would be faster than a SSD (which the other three ahead of me are using for projects).  Perhaps that RAID controller is responsible for some of the blinding speed?

       

      So, Harm, please enlighten us. What the heck have you done? This is the equivalent of Fuel injected engines being left in the dirt by some Atomic powered supercharged monstor under your hood! Quite impressive.

       

      I can look at the list of parts you put into the thing, but I can't (and perhaps many others can't) really determine which components did what?

       

      Edit:  Whoops! I just realized my wife was signed in, not me.  Sorry about that - Steven

       

      Message was edited by: Steven L. Gotz

        • 1. Re: Harm's Monster is FAST
          Harm Millaard Level 7

          Laura, Steven

           

          I tried to put everything in perspective on this page: Planning & Building a NLE System and the following pages (there are lots of them).

          It is not finished yet. I'm still documenting it and trying to get to grips with Win8. What a hassle.!!!!

           

          My results are not optimized yet, they are a simple reflection of where I am now, but some tuning, optimization still needs to be done. I had several background programs, processes still on when I measured these results, but the major distinction is probably the E: drive raid aray that gives a sustained tranfer rate of around 1300 MB/s. Also the 64 GB memory helps. The major drawback is that CS6 is severly handicapped by AME. Exporting by the AME queue takes 37 seconds for the Disk I/O test (PAL DV) but takes only 11 seconds using direct export. That is more than 300% slower, just by the overhead of using AME, as acknowledged by certain Adobe employees.

           

          I expect to shave off several seconds from this timing when I have finished tuning my system. For instance my MPE On figures fluctuated around 1 to 2 seconds, due to measurement errors, but I took them up as 2 seconds.

           

          I will document my tuning tips on the Final Results page but please have some patience. I'm stii working out the irks of Win8.

          • 2. Re: Harm's Monster is FAST
            Jim_Simon Level 8

            the overhead of using AME, as acknowledged by certain Adobe employees.

             

            Have they offered any sort of explanation, or only an acknowledgment?

            • 3. Re: Harm's Monster is FAST
              Jim_Simon Level 8

              the irks of Win8.

               

              Been using it for a while.  What irks you?

              • 4. Re: Harm's Monster is FAST
                Bill Gehrke Adobe Community Professional & MVP

                I would like to comment on Harm's info and add some background info.

                Harm Millaard wrote:

                 

                ......major distinction is probably the E: drive raid aray that gives a sustained transfer rate of around 1300 MB/s. Also the 64 GB memory helps. The major drawback is that CS6 is severely handicapped by AME. Exporting by the AME queue takes 37 seconds for the Disk I/O test (PAL DV) but takes only 11 seconds using direct export. That is more than 300% slower, just by the overhead of using AME......

                 

                 

                 

                The exporting of the Disk I/O directly is taking a six second AVI clip which is repeated it 1092 times and exporting it to one contiguous video.  In the process above there is generated an AVI clip that is ~1 hour in length.or 12.7 GB.  Therefore Harm's 11 seconds to write (in Premiere) a file of 12.7 GB (13005 MB/s) means that Harm's write rate is  1182 MB/sec on that trial as these a whole lot asynchronous mechanical rotating devices and we only have a 1 second clock for timing you can see there lots of room for variability of the scoring.

                 

                I am now using a proposed PPBM6 timeline with about 2-hours of clips and gathering data on it and rather than 11+/-1 seconds we would have 22 +/-1 which halves the variability.

                 

                One interesting experiment I have performed is to convert the clip from AVI to MOV.  Since AVI is native onlly to Microsoft and not on Apple a disk test has to be reinvented for Mac's  Well guess what, initial tests show very similar results on identical hardware platforms.

                • 5. Re: Harm's Monster is FAST
                  Harm Millaard Level 7

                  Apps that you need to uninstall, access rights that need to be adjusted all over the place, a simple inheritance does not work, a profile picture that can't be adjusted, using Windows Defender and Firewall due to lacking Symantec support for Win8, missing printer drivers and Asus utilities, no way (at least at the moment) to bypass the Start page, no easy way to modify the size of the icons on the start page, the problems to clean up the mess Adobe Master Collection makes of the installation (see: http://forums.adobe.com/thread/1085879?tstart=30), the waiting for Win8 certified programs, all stuff like that.

                   

                  The explanation I got is that it is related to the overhead of AME and the Dynamic Link Manager, which I think is pretty lame. I mean the difference between 11 seconds direct export and 37 seconds via the Queue is so huge, that serious optimization of the code is necessary IMO. Note that I have seen this before with the same relative figures on my old i7-920 system. The other thing to note is the serious slow-down of MPEG2 export in CS6 versus CS5 and CS5.5. It sure looks like each new version gets slower in that department. The same timeline in CS5 took 25 seconds on my old system and now takes 50 seconds on a way faster system. I can't explain that. Sure there may be some influence caused by Win8 versus Win7, but not enough to explain this huge difference and I'm not the only one to experience this enormous slow-down.

                  • 6. Re: Harm's Monster is FAST
                    ECBowen Most Valuable Participant

                    The results were with Win 8 then Harm?

                     

                    Eric

                    ADK

                    • 7. Re: Harm's Monster is FAST
                      Harm Millaard Level 7

                      Yes Eric, with Win8 64 Enterprise, build 9200

                      • 8. Re: Harm's Monster is FAST
                        ECBowen Most Valuable Participant

                        Excellent, thanks for the heads up.

                         

                        Eric

                        ADK

                        • 9. Re: Harm's Monster is FAST
                          Jim_Simon Level 8

                          Man, your experience is a bit different than mine, Harm.

                           

                          What apps do you feel you need to uninstall?  I haven't had any permissions issues myself.  I was able to change my profile picture, though.  I can help with that if you want.  I personally would always choose Defender over anything Semantic, so...  I do find the Win7 printer drivers work fine for me.  Have you had issues using those?  I'd be surprised if the same didn't hold true for Asus utilities, Win7 version should work fine.  True, you can't bypass the Start screen, but once you click into the Desktop, the experience is incredibly similar to Win 7.  I never see or use the Start screen except when I first boot up.

                          • 10. Re: Harm's Monster is FAST
                            Harm Millaard Level 7

                            I haven't had any permissions issues myself.

                             

                            Have you tried accessing and modifying any files in the C:\Users\<username>\AppData\Roaming directory  or in C:\Windows\Performance\WinSat\DataStore? You will know what I mean when you try it.

                             

                            I was able to change my profile picture, though.  I can help with that if you want.

                             

                            Look here:

                             

                            Personalize.png

                             

                            It is all greyed out. Nothing to change or modify.

                             

                            I personally would always choose Defender over anything Semantic, so...

                             

                            I much prefer Symantec Endpoint Protection Enterprise over anything Windows, but that may be personal.

                             

                            I do find the Win7 printer drivers work fine for me.  Have you had issues using those?

                             

                            Yes, all those did was put out a lot of scrambled pages, but nothing readable.

                             

                            I'd be surprised if the same didn't hold true for Asus utilities, Win7 version should work fine.

                             

                            They won't install at all, saying that this Windows version is not supported.

                            • 11. Re: Harm's Monster is FAST
                              JEShort01 Level 4

                              Awesome indeed Harm!!! The monster "untuned" as you call it is sure looking great.

                               

                              I'm confused by the hardware listing on PPBM5 with 24 project drives listed, 18TB array size, and total # HDs at 19.

                               

                              Do you have a single RAID 30 with all PPBM5 input and output files on it?

                               

                              What drives did you end up using and how many?

                               

                              How would describe how the monster "feels" for real timeline work vs. your i7-920 rig?

                               

                              Thanks,

                               

                              Jim

                              • 12. Re: Harm's Monster is FAST
                                Harm Millaard Level 7

                                I'm confused by the hardware listing on PPBM5 with 24 project drives listed, 18TB array size, and total # HDs at 19.

                                 

                                So am I. It should read 25 and I have been looking into the MySQL code to discern where the error is, but have not yet found it. It is a calculated field, so my query code must contain an error. It has my priority to solve this. I have the feeling it has to do with the submission form, and the enumerated fields, that only allowed up to 16 disks, not the 24 I currently have.

                                 

                                Found it. The enum structure in the database was the culprit. It has now been corrected.

                                 

                                Do you have a single RAID 30 with all PPBM5 input and output files on it?

                                 

                                Yes. It is a single volume consisting of 3 x (7 disk Raid3) striped to Raid30, so 21 disks, giving me 18 TB net space, plus 3 global hot-spares. It does cost me 6 disks in net space, but it buys me safety in case of disk failure. Even if all three raid3 arrays were to give me trouble at the same time, I would still have the three hot-spares to compensate for that failure.

                                 

                                What drives did you end up using and how many?

                                 

                                Seagate Constellation ES 1 TB Enterprise edition disks, 25 of them for the moment, but possibly I may extend that with 6 more if the need arises. Then again I may also expand on my SSD's from 2 to 4 if funds allow. I have the space and connections internally.

                                 

                                How would describe how the monster "feels" for real timeline work vs. your i7-920 rig?

                                 

                                Awesome, but I have to tune it further to get the results I wanted. I think I can shave off a few seconds here and there.

                                It is a pity that Adobe seems to punish us with each new version with some performance anomalies. If you take my CS5 MPEG2 time (25 seconds) on a much lesser system, my scores could have been:

                                 

                                98 = 37 + 25 + 34 + (39) + 2 or even, without the overhead of AME:

                                 

                                72 = 11 + 25 + 34 + (39) + 2

                                 

                                It would seem likely that this much faster system would be able to improve in the 25 second MPEG2 score, but I have not included that above. The hypthetical assumption would be to expect a score of less than 70 seconds, however.

                                 

                                BTW, I have updated the Top 20 Performers page and the MPE Gains page to show the overall results, like it used to do, but added the same graphs for CS6 machines only when you scroll down.


                                • 13. Re: Harm's Monster is FAST
                                  Harm Millaard Level 7

                                  I just managed to shave off 9 seconds from my score. I tested and retested five runs of each test and usually they were absolutely identical. When there were differences in the scores I took the arithmetic average as an average rounded to the closest integer. The most distinctive figure is my render time, now down to 1 second, but we (at least Bill and I) know that is the measurement accuracy of the Windows system, so what do you do when you have done this test and 7 out of 10 is shows as 1 second and in the remaining cases it is 2 seconds? Well, the average is lower than 1.5 and thus the integer value is 1, which is what I reported. But the remaining three cases were 2 seconds, just so you know. Realistically, the score would be around 1.46 seconds or in some cases 1.52 seconds, but Windows does not give that accuracy.

                                   

                                  I also optimized my disk setup a bit so that took off another few seconds and finally I ended up with a score of 119 seconds. On the Passmark 8 benchmark test I ended with a system score of 6920, also one of the top scores.

                                   

                                  The most disappointing result is the 50 second MPEG2-DVD score, which is 200% slower than my far less capable i7-920 achieved with CS5+. It you look at the MPEG2-DVD graph, you can almost immediately identify the CS6 results by just looking at the double length of the results. Funny (or exactly the opposite ). A clear and demonstrated 200% or more loss in performance with CS6 over previous versions.

                                   

                                  Just to keep you informed.

                                   

                                  PS. By slightly overclocking the video card the MPEG2-DVD decreased to 45 from 50 seconds.