Concerns about PS CS5 scalability

[BernardLanguillier]

Wondering why I keep paying 1.000 US$ every 2 years for this...

http://macperformanceguide.com/Reviews-MacProWestmere-Photoshop-CoresSlower.html

Cheers,
Bernard

p.s.: the said sum is the price of the Creative suite upgrade in my geo

[feppe]

Wondering why I keep paying 1.000 US$ every 2 years for this...

http://macperformanceguide.com/Reviews-MacProWestmere-Photoshop-CoresSlower.html

Cheers,
Bernard

p.s.: the said sum is the price of the Creative suite upgrade in my geo

I sure hope Chris Barrett doesn't read this

[Alan Goldhammer]

Given Adobe's interesting relationship with Apple, maybe this is an intentional design feature; sort of like a stealth effort to get people to move to Windows!

[Eric Myrvaagnes]

It makes me glad to have a lowly 4-core machine (and a PC at that).

Eric

[Wayne Fox]

Interesting read.  Wondering where they got a 3.33 ghz 12 core.  Fastest 12 core I see at the Apple store is 2.93ghz.

I'm having a little bit of a credibility issue with digilloyd, after reading his ranting on adobe's forum and on his site about having to "warm" photoshop up to preallocate memory, because I'm not seeing this at all. So something about his test machine is far different than mine.  From a cold boot, running his tests take about 10% longer the first time through than the second, a far cry from the results he claims and has been arguing with Adobe about in their forums.  Also curious as to the validity of his "test" and the test image.  After all, he only really tests a couple of things ... a resize to a very large document (20,000 pixels wide) and then rotating it a bunch of times.  Does this translate to every task Photoshop does, such as smart sharpen?  I guess logically it might, but the test seems a little limited.

That being said, Adobe has always stated that more than 4 cores on either platform offers little benefit ... to Photoshop.  Digilloyd seema to think it's improper coding by Adobe and maybe he's right ... I'll admit I"m not an engineer but it could easily be many of the tasks in Photoshop just can't be threaded out enough to take advantage of the cores.

However, this would assume the only thing you ever do is photoshop, and in my case I have 10 or 12 apps running at once, often am working in photoshop while Lightroom is working on a couple of tasks, and all these cores do help there.  And my own action on one of my p65 files running on my previous 2008 8 core mac took 55 seconds, and now completes in only 22 seconds.  So maybe turning of some cores might help it, but bottom line it's screaming fast from where I was (and that machine wasn't too bad). 

[BernardLanguillier]

That being said, Adobe has always stated that more than 4 cores on either platform offers little benefit ... to Photoshop.  Digilloyd seema to think it's improper coding by Adobe and maybe he's right ... I'll admit I"m not an engineer but it could easily be many of the tasks in Photoshop just can't be threaded out enough to take advantage of the cores.

However, this would assume the only thing you ever do is photoshop, and in my case I have 10 or 12 apps running at once, often am working in photoshop while Lightroom is working on a couple of tasks, and all these cores do help there.  And my own action on one of my p65 files running on my previous 2008 8 core mac took 55 seconds, and now completes in only 22 seconds.  So maybe turning of some cores might help it, but bottom line it's screaming fast from where I was (and that machine wasn't too bad). 

All valid points Wayne.

My view though is that PS is an image edition platform these days, and that the #1 quality it should have is stability and performance. I have communicated this to Mr. Nack already, I'd rather have his engineers focusing on making PS more scalable instead of working on HDR.

I believe that the effort they have made to move PS CS5 to 64 bits on Mac was significant and shows that Mr. Nack has been listening, but I sure hope that they will continue investing towards more scalability.

As far as Digilloyd's credibility is concerned, well everybody has an agenda of some sort as well as some personal relationship/constraint that make totally objective information reporting difficult. On this one, I believe that all of us with mulit-core machines (even my 3+ years 8 core Mac Pro) know that PS isn't that great at using many cores. Being limited to 4 when the best machines have 24 today and will have 32 by the end of next year isn't acceptable for a high end piece of software.

Cheers,
Bernard

[Wayne Fox]

All valid points Wayne.

My view though is that PS is an image edition platform these days, and that the #1 quality it should have is stability and performance. I have communicated this to Mr. Nack already, I'd rather have his engineers focusing on making PS more scalable instead of working on HDR.

I believe that the effort they have made to move PS CS5 to 64 bits on Mac was significant and shows that Mr. Nack has been listening, but I sure hope that they will continue investing towards more scalability.

As far as Digilloyd's credibility is concerned, well everybody has an agenda of some sort as well as some personal relationship/constraint that make totally objective information reporting difficult. On this one, I believe that all of us with mulit-core machines (even my 3+ years 8 core Mac Pro) know that PS isn't that great at using many cores. Being limited to 4 when the best machines have 24 today and will have 32 by the end of next year isn't acceptable for a high end piece of software.

Cheers,
Bernard

I used an incorrect term in my hasty response.  I'm really not questioning digiloyds credibility, he seems very thorough and knowledgable.  However, the problem I have is his results are nothing close to mine, which means there is something very different about how he sets up his mac.  For example, he believes that Photoshop has a serious problem allocating memory, and has tests which show running the same thing twice results in significant differences in time ... I can't replicate this at all.  And in his recent test where he shows fewer cores is "faster", I can't replicate that either.  Turning off hyperthreading does improve his tests, but reducing cores is never faster on my machine.  For those curious, here is some of my results, running his digilloydmedium tests at 3 interations, all from a reboot, and using the Mac developer processor preference pane.

Standard 12 core/2.93 configuration ... 25-22-23 seconds.
Same, with hyperthreading disabled ... 22-21-21 seconds
No hyperthreading, only 6 cores ........ 26-25-24 seconds
No hyperthreading, only 4 cores ........ 29-27-26 seconds

Note that I can't get any result under 20 seconds, which implies his SSD setup contributes significantly, since I'm using 4 drives in a raid 0 instead. But in my case,
fewer cores is slower, albeit not by much.  One additional test I did was having LR create 50 1:1 previews on phaseone p45 files in the background, a pretty intense task while running Digilloydmedium in the default 12 core  configuration - results 29-27-26, and LR churned merrily along and didn't seem to slow down when I started the PS test. Not bad considering the disk activity and crunching LR is doing.

To your point, the additional cores do not seem to be scaling as one would hope, which I believe is a universal problem with many applications, and not unique to OS X or Win 7.  Whether this is just a limitation of threading operations or not I can't say ... hopefully not hopefully Adobe is working on ways to leverage more cores.

But the speed gain I'm seeing from my previous 2 year on 8 core mac is significant and welcome.

[ErikKaffehr]

Wayne,

Thanks for sharing good info. My impression is that you really need to have something like a 30% difference for before it starts to be significant. What Lloyd says that there is little gain having 12 CPU-s over 6 CPU-s for Photoshop related work. The main reason for this is probably that a large part of the underlying code is not optimized for multiprocessing (fully reentrant). So Adobe can do some smarts on the front end, but soon enough they get into code that needs exclusive access to variables. The only way out of this is optimizing all the underlying code for parallel processing.

So Lloyd says, buy quad or hexa core with the highest clock rate, because Photoshop won't scale beyond four CPU-s.

Part of the problem is that the PC industry was focused on MHz and now GHz for a long time. The Pentium 4 was intended to go past 10 GHz, but than it was realized that power consumption was going up and cooling a 10 GHz CPU was not feasible. So Intel went back to the Pentium Pro/Pentium 3 design but started adding more cores to each chip. Now we suddenly had parallel processing on the desktop but no parallel software and no experience in developing parallel software for the desktop. Parallel computing is nothing new to supercomputing, like weather forecast, but very new to office type of software.

Best regards
Erik

[Wayne Fox]

Wayne,

Thanks for sharing good info. My impression is that you really need to have something like a 30% difference for before it starts to be significant. What Lloyd says that there is little gain having 12 CPU-s over 6 CPU-s for Photoshop related work.

Actually what he is saying is 12 cores is slower than 6 cores which means 12 cores degrades PS performance.

  So while I would agree that going from 6 to 12 cores doesn't scale like one would hope (if the machine is running only Photoshop CS5), in reality it does go faster with 12 cores than 6 for me.

So does buying a 12 core machine actually get you a slower workstation?  I'm not seeing it, although he claims it will unless you disable some of the cores.

[Wayne Fox]

For those curious, here is some of my results, running his digilloydmedium tests at 3 interations, all from a reboot, and using the Mac developer processor preference pane.

Standard 12 core/2.93 configuration ... 25-22-23 seconds.
Same, with hyperthreading disabled ... 22-21-21 seconds
No hyperthreading, only 6 cores ........ 26-25-24 seconds
No hyperthreading, only 4 cores ........ 29-27-26 seconds

For the curious that stumble onto this thread, I thought I'd add another couple of tests I did.  Obviously running multiple apps can make use of the additional processing power.

Running digilloyd medium in Photoshop front application, LR creating 1:1 previews in background
12 core/2.93ghz, hyperthreading on. . . 29-27-26
4 cores /2.93ghz, hyperthreading on...  41-40-40

digilloyd medium in Photoshop front application, LR creating 1:1 previews and Handbrake ripping a DVD in background
12 Cores, HT on  34-34-32
4 cores, HT on .. 54-57-55

[Farmer]

If I had to choose between DigiLloyd or the likes of Jerry Harris and Chris Cox and Eric Chan as to who I believe about what's going on under the hood, I'll choose Chris and Jerry and Eric every time.

Obviously there's more work to be done, but Adobe has said they're continuing to work to improve these things.

Sure, Bernard wants more speed and not HDR, but a lot of people wanted better HDR.  None of us are likely to ever be 100% satisfied with the decisions, but that's life.

I'm sure that DigiLloyd has good intentions, but the tests seem far too limited and the results questionable (look at Wayne's results, for example).  There's a lot of noise being made by someone who doesn't have access to the code to know what's actually happening and who isn't testing all the various modules but then tries to give a general opinion. 

That's not good science.

[ErikKaffehr]

Hi,

To begin with I would say that Adobe puts a lot of effort in improving solutions between versions. HDR starts to be quite usable for instance, in my view. There will always be some software that is better for a given purpose but Photoshop starts to be good enough and has workflow benefits.

Regarding the "Diglloyd" articles I would say that what he is saying is that "save your money and buy single CPU and high clock rate". According to him single CPU with highest clock rate on present generation Intel CPUs is the "sweet spot" for most applications.

Some of Lloyd's advice may be a bit strange to me, to begin with Wayne Fox doesn't seem to be able to reproduce the effect of disabling HT, the other issue I may have that he is a bit over obsessed with disk I/O. When I check "Performance Monitor" on my Mac I very seldom see significant disk activity, except when copying files. Obviously, having short seek times will help many applications even if they are not bandwidth limited on disk I/O. Using SDDs for scratch disks may be very smart.

I may also suggest that we have different scenarios:

1) Photoshop works fine, be happy!

2) Photoshop is slow with your files.
- Learn to use activity monitor
- Check CPU utilization, if it's low neither more or faster CPUs will help
- Check memory utilization if memory usage is very high you may need to add more memory (Lloyd has info on that, too).
- Check I/O activity

3) If the computer is very sluggish you are probably low on memory

Best regards
Erik

Sure, Bernard wants more speed and not HDR, but a lot of people wanted better HDR.  None of us are likely to ever be 100% satisfied with the decisions, but that's life.

That's not good science.

[kers]

To begin with I must say that Photoshop CS5 gives the best speed improvement of all the upgrades so far.

Opening and saving is the bottleneck i think, using 1 core!
- we need an new PSD-format that can be saved in parts - modular- each part saying if it has been changed or not.
I have noticed that is uses more cores than ever on some assignments- the best being the image processor. It uses all the 8 cores and is the fastest batch RAW to Tiff solution I have.
Here it does about 12 D3x files Nef14bit to tiff 16 bit uncompressed in 1 minute)
The warm-up speed issue detected by Diglloyd can be confirmed here on my machine. It seems the allocation of ram takes place during the work and not before so the first time things that need more Ram takes longer. I can imagine that at Adobe they find it not that important, but the way they responded to Diglloyd was not very client -friendly - to say the least- nor very adobe promotional either. They should be happy with people that show the shortcomings of their software.
I must say Diglloyd does a very good job testing all this stuff and finding things i never really thought of. Like the speed issue with the 12 core machine.

Wayne-  maybe i have an old ( different) digllloyd medium test but my old 2008macpro- 8 core- 16 gig machine seems to do it in 12 seconds? ( have an OSX-SSD that runs in 64 bit giving photoshop 12 gig ram)
do you have the blocksize  at 1024? 

cheers PK

[Wayne Fox]

Wayne-  maybe i have an old ( different) digllloyd medium test but my old 2008macpro- 8 core- 16 gig machine seems to do it in 12 seconds? ( have an OSX-SSD that runs in 64 bit giving photoshop 12 gig ram)
do you have the blocksize  at 1024? 

cheers PK

Yes, tile size is set to 1024.  12 seconds is amazingly fast ... faster than digilloyd himself has ever been able to achieve with any setup.  You were running the test from the applescript and not just on a random file you opened?  My 2008 dual quad core takes about 55 seconds to run digilloyd medium.

[JeffKohn]

I think expecting scalability up to 10-12 processors (or more) is not necessarily realistic, at least not for all operations (or even most of them). There are many "problems" that just can't be broken down enough to support that sort of parallel processing.

Of course, some operations are not CPU bound, but rather memory- or disk-bound. So just throwing more CPU's at a configuration may not help anything. I think that for most software that people are using today, 12 cores amounts to overkill, especially if those cores are hyper-threaded. Even if you have enough background tasks to utilize all those cores, other parts of the system are bound to create bottlenecks, and in fact may slow down performance of the foreground task.

Now, if Lloyd's assertion that PS is actually slower on 12-proc system than an otherwise identical 4-processor is valid, that would certainly be troubling. It would seem to indicate that PS isn't just limiting itself to one (or two or four or whatever) processors, but that it is attempting to use all processors and doing so very poorly. There's no real excuse for that if true, but given Wayne's inability to reproduce those findings it looks like there's more to this than just processor count, and something else about Lloyd's scenario with his configuration needs to be considered.

[kers]

Yes, tile size is set to 1024.  12 seconds is amazingly fast ... faster than digilloyd himself has ever been able to achieve with any setup.  You were running the test from the applescript and not just on a random file you opened?  My 2008 dual quad core takes about 55 seconds to run digilloyd medium.

Ha-  then there must be some apple-pear problem here ! 
I just start of with a random image that is scaled in the action to 20000 pixel

I am sure I can't be faster then Diglloyd....

[Wayne Fox]

Ha-  then there must be some apple-pear problem here ! 
I just start of with a random image that is scaled in the action to 20000 pixel

I am sure I can't be faster then Diglloyd....

The applescript will create the test file and ask you to save it.  Part of the test is the actual  resize to 20,000 pixels from the test file, which means your test skips that step.

The script also asks which test you want to run, and how many iterations. It runs the actions in photoshop for you, and  then tells you the results of each iteration as well as the total time taken.

The script is included in the zip file when you download the actions.

[tived]

I think you will find that I/O is the single biggest bottleneck in your computer (be it PC or Mac)
the best bang for you buck is adding SSD's and then memory, then CPU's.

Henrik

Hi,

To begin with I would say that Adobe puts a lot of effort in improving solutions between versions. HDR starts to be quite usable for instance, in my view. There will always be some software that is better for a given purpose but Photoshop starts to be good enough and has workflow benefits.

Regarding the "Diglloyd" articles I would say that what he is saying is that "save your money and buy single CPU and high clock rate". According to him single CPU with highest clock rate on present generation Intel CPUs is the "sweet spot" for most applications.

Some of Lloyd's advice may be a bit strange to me, to begin with Wayne Fox doesn't seem to be able to reproduce the effect of disabling HT, the other issue I may have that he is a bit over obsessed with disk I/O. When I check "Performance Monitor" on my Mac I very seldom see significant disk activity, except when copying files. Obviously, having short seek times will help many applications even if they are not bandwidth limited on disk I/O. Using SDDs for scratch disks may be very smart.

I may also suggest that we have different scenarios:

1) Photoshop works fine, be happy!

2) Photoshop is slow with your files.
- Learn to use activity monitor
- Check CPU utilization, if it's low neither more or faster CPUs will help
- Check memory utilization if memory usage is very high you may need to add more memory (Lloyd has info on that, too).
- Check I/O activity

3) If the computer is very sluggish you are probably low on memory

Best regards
Erik

[Wayne Fox]

the best bang for you buck is adding SSD's

Henrik

Personally I've debated the SSD route, but having a hard time understanding how using an SSD can really help all that much (except perhaps app launch speed and first time you do some tasks) if  you have enough RAM.  Currently Photoshop rarely shows less than 100% efficiency meaning it isn't using the scratch disk.  Wondering if the SSD might be related to diglloyd's tests because of virtual memory issues with the OS.

Maybe I'll break down and load one up ... but then there's the whole 1SSD or 2 stripe 0 SSD thing ... or 2 non stripped SSD's after reading a few threads I'm not sure which one is really the best.

[kers]

Personally I've debated the SSD route, ....

I agree with Henrik.
i replaced my system OSX disk for a SSD and it works very fast overall. Every program starts op with one jump- the system start up within 15 seconds. (the clock makes 6-7 rounds)
What i did not do yet is replacing the work disks ( 2 striped raid) for two SSD's  .
I find that still too expensive.