[Shaktai]

Well, my computers have been moved back over to Einstein to help the team out there, but out of 83 work units completed with the alpha optimized apps, 81 have validated, 1 is still pending and 1 errored on download (not on execution). I would say that is a successful app.

[rick]

Hi, I'm the other developer that Alex keeps bitching about. We're going to be hosting stuff on my web space in the future so please redirect your attentions to http://writhe.org.uk/seti@home/ for new clients.

It's a bit sparse at the moment but I'll get round to adding version numbers and list of changes at some point. The source code is also up for those who want to know where the Magic comes from.

Also, the clients are now (purposefully) split into just G4 and G5 clients (no G4 7400 / 7450 split). If this proves to be significantly slower I'll try mucking about with the compiler some more. It seems that in some cases specifying an exact architecture (e.g. 7450) to the compiler is slower than just a generic architecture (e.g G4). Pfff, compilers...

The new version (alpha 4) gets the run time down to just under 2 hours on my 1 GHz PowerBook on the test work unit. For comparison, the alpha 3 client was around 2½ hours.



I admit that I haven't done as much testing of alpha 4 compared to previous versions so if anything goes haywire just resort to the previous versions posted by Alex. Send any error reports my way because it will be my fault, either through this forum or e-mail (details are on the site). I'm going to be away in London for the next few days so if anything is wrong it won't get fixed until Monday or Tuesday.

I suppose if you can't wait then you can try abusing Alex.



As for the long term state of optimisation, there are probably a few more tweaks that we can squeeze out of it. The main problem is that squeezing more speed out of it requires much more work than previous optimisations. Whereas before we could just do something like use Apple's optimised code and cut the time from 8 hours to 4 hours, now we're resorting to a larger number of smaller optimisations and getting down to the bare metal.

There is still a bit more to be done for those of you on Mac OS X 10.3 as the backwards compatibility routines haven't been optimised yet. This is mainly grunt work (relatively easy, just a lot of testing) and we're focusing our efforts on optimisations that will help everyone at the moment, but yeah, eventually the 10.3 clients should get a fair bit faster.



As for all the requests about graphics support and a single client with CPU detection, all I can say is "yeah, eventually".

There's a load of stuff that we'd like to get implemented at some point. We are but 2! Hopefully our code will get integrated into the main SETI@home version after it's been cleaned up a fair bit. Although, I'll be honest and say our code is a damn sight better than what we had to work with.

Have fun with the new client and keep us posted.



r i c k

[alexkan]

mikkyo or Shaktai: Any chance that we could get a plug on the Team MacNN website, now that we have a respectable-looking webpage up? I suspect that a couple of the biggest crunchers probably stand to gain a lot from using the new clients.

[reader50]

I'll do it later tonight if need be, but mikkyo or Shaktai would be better choices. Unlike me, they've actually used the software in question. I'll be setting up my very first BOINC stuff this weekend.

If you guys need client downloads hosted, or links from our downloads pages, let me know. We put up links for javalizard's client, for example.

[alexkan]

Originally Posted by reader50

I'll do it later tonight if need be, but mikkyo or Shaktai would be better choices. Unlike me, they've actually used the software in question. I'll be setting up my very first BOINC stuff this weekend.

If you guys need client downloads hosted, or links from our downloads pages, let me know. We put up links for javalizard's client, for example.

Hey, if you want to link us off our your download pages, that would be great, too.

[reader50]

alexkan, rick, several questions.

Your new page does not list alpha-3, in case anyone has trouble with alpha-4. Did you want the old locations linked to as well, or just the new page?

I'll link to the page(s) of course. However, if you'd like to save the latest version with a fixed name, I could put a direct download link to it as well. Something that won't break when alpha-5 come out.

Finally, did you want your stuff on our regular download page, or the beta downloads for now? The fact that it's called an alpha suggests the beta page.

Let me know how you want it done.

[alexkan]

Originally Posted by reader50

alexkan, rick, several questions.

Your new page does not list alpha-3, in case anyone has trouble with alpha-4. Did you want the old locations linked to as well, or just the new page?

I'll link to the page(s) of course. However, if you'd like to save the latest version with a fixed name, I could put a direct download link to it as well. Something that won't break when alpha-5 come out.

Finally, did you want your stuff on our regular download page, or the beta downloads for now? The fact that it's called an alpha suggests the beta page.

Let me know how you want it done.

Yeah, I think linking to alpha-3 as backup purposes is probably a good idea. The links are all of the form http://inst.eecs.berkeley.edu/~alexkan/seti/sah-4.18.ppc$NUMBER-vdsp-a3.tar.gz, where $NUMBER is either 7400, 7450, or 970. I'll make sure I leave them up, but I'll need to move them when I graduate and my account gets deleted. Hopefully no one will need alpha-3 by then.

I think it might be good to direct people to the page, since it contains useful information about where to report issues. Maybe once we're sure that we're done with it, we can make a direct link, but for now, I still see a version number increment or two.

In my opinion, alphas 1-3 have seen enough action to be considered ready to go (Shaktai can probably attest to this), but alpha-4 is still technically an unproven quantity, although from what I've seen, I have no reason to believe it'd be any less reliable than the other alphas. It would probably be more fair to put it on the beta page, but considering the amount of speedup that this client produces over javalizard's client, I think it's probably worth at least mentioning on the regular download page.

[reader50]

It is up. Main page, because from the thread responses, it seems to be stable. And oh so fast.
http://members.dslextreme.com/~reade...eam/boinc.html

I'm assuming you will link to the alpha3 clients from that page, so I didn't put a separate link. Also, I alternated your names so as not to play favorites.

Your names are listed exactly as printed in the forums. If you'd like something different, let me know. ie: Capitalization, custom titles "SETI Compile Masters, Inc.", links to favorite illegal download sites, etc. Uh, on 2nd thought, scratch that last service.

Let me know if it's all kosher. If so, we can post a news blurb.

[alexkan]

Originally Posted by reader50

It is up. Main page, because from the thread responses, it seems to be stable. And oh so fast.
http://members.dslextreme.com/~reade...eam/boinc.html

I'm assuming you will link to the alpha3 clients from that page, so I didn't put a separate link. Also, I alternated your names so as not to play favorites.

Your names are listed exactly as printed in the forums. If you'd like something different, let me know. ie: Capitalization, custom titles "SETI Compile Masters, Inc.", links to favorite illegal download sites, etc. Uh, on 2nd thought, scratch that last service.

Let me know if it's all kosher. If so, we can post a news blurb.

Yes, I think "oh so fast" characterizes the new client quite well.

I realize that I'm speaking for Rick, since he's probably been sleeping for the past couple hours, so I hope he won't mind, but since both our forum names are based off our real names, could you put our names as Alex Kan and Rick Berry?

Rick's the only one who can touch the web page, so links to alpha-3 on the webpage will have to wait. I can restore them to the first post of this thread, though, at some point.

[reader50]

Names edited. Also, news story posted.
http://team.macnn.com

Hope you like my editorial style. It can sometimes be a mistake to post late at night.

[mikkyo]

The G5 version of alpha-4 was compiled with arch ppc7400, not ppc970 as I would expect.
Is that intentional?
$ lipo -info setiathome-4.18-applevdsp-tiger-G5-alpha-4
Non-fat file: setiathome-4.18-applevdsp-tiger-G5-alpha-4 is architecture: ppc7400

[alexkan]

Originally Posted by mikkyo

The G5 version of alpha-4 was compiled with arch ppc7400, not ppc970 as I would expect.
Is that intentional?
$ lipo -info setiathome-4.18-applevdsp-tiger-G5-alpha-4
Non-fat file: setiathome-4.18-applevdsp-tiger-G5-alpha-4 is architecture: ppc7400

As far as I know, 970 compiles (like the ones I'm currently testing, which I just ran lipo on) should be arch ppc970. This could be a packaging mistake--I'd post new binaries just to be sure, but not only is Rick the only one who can modify the website, but I'm in the middle of testing some more optimizations and it'd be a bad thing to post what I'm working on right now.

Thanks for bringing this to our attention.

[beadman]

Alex:
Thanks for your hard work getting us the new optimized clients! Is there any chance you'll be trying to do some optimizations for the G3 machines any time soon? Some of us still have an old imac or other G3 quietly plodding along with OS 10.3.9 and SETI using the default BOINC and SETI clients...

Thanks!
beadman

[chboss]

Alex, Rick

Thank you for bringing highly optimized Seti clients to the Mac community!

My own experiments to produce a faster Seti clients resulted in a modest gain of only 20% compared to the original Berkeley client. Nohwere near the fantastic gain of your clients.

I am currently running the 7450 Alpha-3 on three boxes: Mac mini 1.25GHz, G4 1GHz (Sonnet upgrade) and a PowerBook 1GHz all running "Tiger" 10.4.2. The Mac mini and the PowerMac run without problems. All results validated until now.

The Power Book showed some problems I've never seen before with the original client or my own clients... It was not able to finish some work units and stalled. It only continued after restarting the client. I have just installed the G4 Alpha-4 client to see if it solves this problem. Any idea?

A small request:
Is it possible for you to add a release note file with some information regarding the changes, fixes, improvements for each version? This would help to see if certain problems are already solved in the newest version.

Keep up the great work!

[Karl Schimanek]

I'm now testing the alpha4 G4 version.
So far, it seems that the alpha4 is as slower as the alpha3
CPU: MPC7450.

Is the G4 core optimised for the 7450 or the 7400 AltiVec-Unit?

[alexkan]

Originally Posted by Karl Schimanek

I'm now testing the alpha4 G4 version.
So far, it seems that the alpha4 is as slower as the alpha3
CPU: MPC7450.

Is the G4 core optimised for the 7450 or the 7400 AltiVec-Unit?

As far as I know, it probably won't be worth it to pursue heavy optimization for G3s, since these optimized clients leverage the Altivec instruction set pretty heavily, which is really the only reason that we're able to realize these sorts of gains. Sorry

With regard to why alpha-4 might be slower than alpha-3--alpha-4 is based off of a totally new build system, and actually doesn't have the same set of modifications made to it as alpha-3. (Rick and I have been working off of our own diverging sets of source code, mailing each other changes as they've been made. No CVS here. ) There are a couple changes that I'm still moving over to the new (Xcode) build system, so keep your eyes peeled for alpha-5.

I'm not sure exactly which G4 architecture the G4 core is optimized for--Xcode is fairly vague on this topic. If it proves worthwhile, we could bring back separate 7400 and 7450 compiles in alpha-5. Also, cache streaming instructions are turned off in alpha-4, which incurs a decent performance hit on the chirp function. (Incidentally, alpha-2 didn't have cache streaming instructions either, but the rest of the client was so much faster than alpha-1 that I doubt anyone noticed.)

About release notes...I hardly remember which changes were made between the alphas myself. I could go back and look at the emails that Rick and I have exchanged, but in general, the changes are just minor tweaks to functions, with a couple function rewrites here and there.

[Karl Schimanek]

Oh sorry, the new core is faster

alpha3: 12.200 seconds
alpha4: 10.900 seconds

MPC7450 733MHz

One question: How many people are working on this optimized cores?

[alexkan]

Originally Posted by Karl Schimanek

Oh sorry, the new core is faster

alpha3: 12.200 seconds
alpha4: 10.900 seconds

MPC7450 733MHz

One question: How many people are working on this optimised cores?

How many results are these times based off of? We need to make sure that we're getting reasonable averages, after all.

Also, it's just me and Rick working on this for now. So yeah, two people.

[Shaktai]

So, I go on the road for a few days (work) and what happens, the forum goes crazy. Won't be able to test alpha4 until I get back. Will try it though with the reference unit.

[Karl Schimanek]

Originally Posted by alexkan

How many results are these times based off of? We need to make sure that we're getting reasonable averages, after all.

Also, it's just me and Rick working on this for now. So yeah, two people.

Okay. Four results, all valid. Longest run 11.150 seconds. An advantage of 17 minutes (versus the longest alpha3).

Karl

P.S. And you should maybe post a message on http://www.simdtech.org/altivec too about the AltiVec core. There are many AltiVec programmers, too.

[chboss]

Here are some timings over several WU's

Running Alpha-4 on 10.4.2:
Mac mini 1.25GHz: 6200s

Running Alpha-3 on 10.4.2:
G4 1GHz (Sonnet upgrade): 8600s

Running Alpha-4 on 10.3.9:
Imac 1.6GHz: 4800s (a big improvement over Alpha-3 used to be around 11000s)

I am still checking what is wrong with my powerbook...

[alexkan]

Originally Posted by chboss

Here are some timings over several WU's

Running Alpha-4 on 10.4.2:
Mac mini 1.25GHz: 6200s

Running Alpha-3 on 10.4.2:
G4 1GHz (Sonnet upgrade): 8600s

Running Alpha-4 on 10.3.9:
Imac 1.6GHz: 4800s (a big improvement over Alpha-3 used to be around 11000s)

I am still checking what is wrong with my powerbook...

Thanks for the timings! Incidentally, alpha-2 and alpha-3 were never released for 10.3.9, so until Rick figured out the Xcode build system, Panther systems weren't reaping the big performance gain between alphas 1 and 2, which is probably what you're seeing on the iMac. You might also want to make sure your processor performance is set to Highest rather than Automatic, since you're presumably running the CPU full-time, anyway.

Also, thanks for posting on the main SETI@home forums on our behalf! I don't have an account on those forums, and AFAIK neither does Rick, so it's nice to see our clients go out into larger circulation. The dual 2.5 in that thread is the fastest machine I know of that's running the clients right now, so I'm looking forward to seeing what kind of numbers it'll put up.

(On a somewhat-related note, Rick's the only one from England...I just happened to see a post he made to a mailing list and offered to give him a hand.)

[chboss]

Hi Alex

Well I kind of rushed ahead with that posting without asking
There are certainly many Mac users that like that kind of performance gain.
I am happy you don't mind.

Sorry for confusing your location. Because Ricks post I thought you were both located in England...
As I can see now you live in a complete different time zone.

Some people confuse me with the real authors of the clients since I posted the link.
Maybe it would be a good idea to add your names to your website.

Setting the processor to full speed is only possible on my PowerBook the Imac seems not to offer this switch. But I will check this again.

Your clients convert the Mac's into real rockets when it comes to Seti crunching

[alexkan]

Well, the website won't be updated until Monday or Tuesday to reflect this since Rick is out of town, but I've got some new clients for you guys to test. (I know, I know, it seems like alpha-4 just came out.) This client should give about a 10% improvement over alpha-4 due to some nifty features of the vDSP libraries that I finally figured out how to use properly. This means that the Gaussian-fitting functions (for those of you who have perused our source code) are almost completely vectorized now. Since the source code is on Rick's website, you might not get to see these changes until he updates the rest of the website.

I enjoy the feeling of squeezing out performance gains when I had thought that there wasn't much room left for improvement. We'll have to see how much longer this lasts.

Without further ado, though, here are the new clients:

G4: http://inst.eecs.berkeley.edu/~alexk...home-G4-a5.zip
G5: http://inst.eecs.berkeley.edu/~alexk...home-G5-a5.zip
(Note: these compiles are currently Tiger-only)

Enjoy, and be sure to report your work unit times!

Chris, can you post a notification to the SETI boards so that people know about this update?

(Also, fun fact for you all: the client spends about 80% of a work unit's processing time in Apple's vDSP functions, so that should be an indication of how we're running headlong into Amdahl's law.)

[chboss]

Alex

I guess these clients are for "Tiger" only?
I will give them a run immediately on my G4's.
Once I know about the OS version I will post to the seti boards.

TIme update:
Powerbook G4 1GHz using Alpha-4 on Tiger: 8300s
It seems something was messed up in the OS... never saw that before on one of my Macs
After a reinstall it is running fine.

[electrician1972]

Under the Alpha-4, here's some timing info
Mac Mini 1.42 GHz - Over the scope of 23 results: Lowest time 6004.93s, Highest time 7074.46s, 2 real short units at 114s and 149s
Mac Mini 1.42 GHz #2 - Over the scope of 9 results: Lowest time 6005.31s, Highest time 6657.30s, 1 real short unit at 102s, and 2 units with errors. Also had 2 units with "Client error". The log showed one of the units had reported an "unrecoverable error". Unfortunately, due to pilot error, I lost that particular log so I can't give any more detailed info.
iBook 1GHz - Over the scope of 8 results: Lowest time 7606.45s, Highest time 8343.22s, 3 real short units at 35s, 66s and 39s.

All 3 computers are running 10.4.2. Crunch harder I say. Clearing the queues for Alpha 5. Thanks Alex...

[Karl Schimanek]

And you should add "AltiVec optimized" or something like that, in the stderr out.

Karl

[alexkan]

Originally Posted by chboss

Alex

I guess these clients are for "Tiger" only?
I will give them a run immediately on my G4's.
Once I know about the OS version I will post to the seti boards.

TIme update:
Powerbook G4 1GHz using Alpha-4 on Tiger: 8300s
It seems something was messed up in the OS... never saw that before on one of my Macs
After a reinstall it is running fine.

Yes, the new compiles are Tiger-only. Since we need to rewrite certain vDSP functions for backwards compatibility, I need to make sure that all the pieces are in place for the Panther compiles.

And yes, dumping information to stderr is probably a good idea...next version, maybe? (Or maybe a revised alpha-5, depending on how things come along.)

[chboss]

OK, I did post the links over at the Seti boards.

The stderr idea is excellent, it allows to identify if the result was achieved with your client.
You should also add some version information.

Here the first results with Alpha-5:
G4 1GHz Powerbook: 7700s (8300s with Alpha-4)
G4 1GHz Sonnet: 6900s (7800s with Alpha-4)

I am using theMacnn superbench Boinc client which is claiming around 23 credits per WU with Alpha-5.
If you improve further we will need a higher optimized boinc client to claim enough credits...

Let me know if you need more information.

[Karl Schimanek]

A higher optimized Boinc client would be excellent, too

Karl

[alexkan]

Originally Posted by Karl Schimanek

A higher optimized Boinc client would be excellent, too

Karl

I know that most of you would probably like optimized BOINC clients, but sadly, there isn't as much latitude for optimization in the BOINC benchmarks. Most importantly, there's no possibility of Altivec optimization at all, since the floating-point benchmark uses double-precision floating point only (Altivec only operates in single-precision). The best we can probably hope for is platform-specific compiles like what are already offered through the MacNN resource pages.

[rick]

Originally Posted by mikkyo

The G5 version of alpha-4 was compiled with arch ppc7400, not ppc970 as I would expect.
Is that intentional?
$ lipo -info setiathome-4.18-applevdsp-tiger-G5-alpha-4
Non-fat file: setiathome-4.18-applevdsp-tiger-G5-alpha-4 is architecture: ppc7400

Originally Posted by alexkan

As far as I know, 970 compiles (like the ones I'm currently testing, which I just ran lipo on) should be arch ppc970.

This is okay. The code is still optimised for G4 / G5 and should run without any performance differences. I assume it was tagged as "ppc7400" because that's what my PowerBook is (well 7450, but 7400 is just the lowest common denominator). I believe the operating system loads the code that is the closest match to the hardware, otherwise it settles for whatever code is available so it doesn't make a difference in practice.

And yeah, thanks for pointing this out, it should be fixed in the next build I put on the site. I'm currently figuring out if it's possible to build binaries with support for specific sub-architectures (7400, 7450, 970) which will require different compiler settings for each.

Originally Posted by Karl Schimanek

I'm now testing the alpha4 G4 version.
So far, it seems that the alpha4 is as slower as the alpha3

Originally Posted by Karl Schimanek

Oh sorry, the new core is faster

Don't scare me like that. You absolute bastard.

Originally Posted by chboss

A small request:
Is it possible for you to add a release note file with some information regarding the changes, fixes, improvements for each version? This would help to see if certain problems are already solved in the newest version.

Will do. Any other suggestions for the site are appreciated, I realise it's a bit terse at the moment.

Originally Posted by Karl Schimanek

And you should add "AltiVec optimized" or something like that, in the stderr out.

Originally Posted by chboss

The stderr idea is excellent, it allows to identify if the result was achieved with your client.
You should also add some version information.

Yeah, what would you guys find useful to be printed? Current version number ("alpha 5") etc.? I don't know if we'll be able to print the work unit name / number because I think the BOINC client essentially just passes SETI a blob of data. I'll investigate.

Timing information would probably be good as well, I'd like to keep it minimal though. Does anyone know if the stderr file gets overwritten for each work unit or just appended to? Alternatively, we might be able to use the BOINC interface and print to it's log file which would seem to be a nicer solution. Is there some reason that stderr would be better?

[chboss]

Rick

What I would like to see in the stderr is:
1. Version information
2. Platform
3. Optimizations
4. The name of the Auhors / Compilers


The Stderr can be checked for each result.
So you can easily find out if the result was calculated with your client and also with which verison.

Here is an example of Ned's AMD64 Linux client:
Result page with customized STDERR

So far my 2 cents...

[mikkyo]

If you build alpha 4 G4 with -arch ppc7400 and G5 with -arch ppc970, you can lipo them together and just have one client that works on G4/G5

[garrett]

These clients are very impressive! I've completed a bunch of work units using Alpha-4 and Alpha-5. So far, no validation problems. Here are times (old times are with javalizard's optimized client):

Dual 2.0 G5 / Tiger / Alpha-5:
Between 3200s and 5100 s, usually about 4400 s
(old time was usu. 11000 s)

G4 1GHz (orig. 400MHz Sawtooth) / Panther / Alpha-4:
8500 s
(old time was usu. 26000 s)

PowerBook G4 400MHz / Panther / Alpha-4:
Between 10,500s and 12,000 s
(old time usu. 40,000 - 50,000 s)


One other thing... I have a question about Alpha-5. Looking at process info. in Activity Monitor, I noticed something. As the client processes a work unit, its memory usage keeps increasing. It starts out at the normal 25 MB or so, but if left running, it can use 130 MB or more by the time it gets done with the workunit. (If you stop the client and restart it, memory usage resets.) I haven't noticed this behavior on my other machines running Panther / Alpha-4. Is there a memory leak or something, or is it simply normal behavior for my setup? (Dual 2.0 G5, 10.4.2, 1.5GB RAM, BOINC 4.44, Alpha-5)
Sorry if I'm being clueless here. I'm not a computer expert by any means.

[alexkan]

Originally Posted by garrett

One other thing... I have a question about Alpha-5. Looking at process info. in Activity Monitor, I noticed something. As the client processes a work unit, its memory usage keeps increasing. It starts out at the normal 25 MB or so, but if left running, it can use 130 MB or more by the time it gets done with the workunit. (If you stop the client and restart it, memory usage resets.) I haven't noticed this behavior on my other machines running Panther / Alpha-4. Is there a memory leak or something, or is it simply normal behavior for my setup? (Dual 2.0 G5, 10.4.2, 1.5GB RAM, BOINC 4.44, Alpha-5)
Sorry if I'm being clueless here. I'm not a computer expert by any means.

I've actually noticed this during the course of testing. We allocate and deallocate a fair amount of scratch space while the worker is running, and were we to have a full-blown memory leak, we would actually exhaust the address space while we're running (which actually happened during one of my tests when I forgot to free one region of scratch space. MallocDebug reports that there aren't any leaks in the client, so I'm not totally sure what's going on. I've sent Rick the changes I made between alpha-4 and alpha-5, so hopefully we'll be able to see if his compiles reproduce the problem.

Thanks for reporting your run times. Can you link to your SETI results page so we can follow the client in the future?

[garrett]

Originally Posted by alexkan

MallocDebug reports that there aren't any leaks in the client, so I'm not totally sure what's going on. I've sent Rick the changes I made between alpha-4 and alpha-5, so hopefully we'll be able to see if his compiles reproduce the problem.

I should mention that since my post I haven't noticed it so dramatically, so it seems to be variable... today it's been staying around 28 MB, other times up to about 50 or 60.

Originally Posted by alexkan

Thanks for reporting your run times. Can you link to your SETI results page so we can follow the client in the future?

Sure. The Dual 2.0 G5 is here:
http://setiathome.berkeley.edu/resul...?hostid=848301


If you care about the others (old dogs, but these clients have given them new life!), the G4 1 GHz CPU-Upgrade Sawtooth is here:
http://setiathome.berkeley.edu/resul...hostid=1209956

and the PowerBook G4 400 MHz is here:
http://setiathome.berkeley.edu/resul...?hostid=211261

[BTBlomberg]

I just have to say thank you to Alex and Rick. I have been using the Javalizard version on my G4s for quite some time now and although it gave ma a boost it was marginal. This is on a PowerBook G4 500 and a PowerMac G4 Dual 866. I only saw a 10% increase over the Seti release. I had looked at compiling a better version but the learning curve was great and the time I have small. So I really appreciate your work.

I installed the Alpha-4 G4 Version less than 24 hours ago and have see these results:

PowerBook G4 500 - MacOS 10.4.2

Time with Javalizard version 4.30
Average of last 5 results: 41,052 Sec.
Average Claimed Credit: 80.16
Average Granted Credit: 23.74

With Alpha-4 G4
Average of last 4 results: 10,049 Sec.
Average Claimed Credit: 19.64
Average Granted Credit: 22.21*
*With only 2 of 4 results validated (Granted)

The reason I added the Credits Claimed and Granted should be obvious. As you can see the Javalizard client was very inaccurate in it's claimed credits and actual granted credits with your new version is in line with what the old JL client got. So, the real benefit is it's about 4 times as fast for me. The same results of 3 or 4 days of work in less than a day.

I will give feedback on the Dual 866 G4 when I get info from it. I put the files in the Boinc but did not restart Seti to finish workunits in progress thinking it Alpha-4 would take over when the JL ones were done and I guess I was wrong. Have a WU at 90 % on one Processor and 20% on the other so when the one at 90% is done I will switch it over.

I always was frustrated by my Pentium 3 550 beating out my G4 500 by an recent average credit of 72.32 to 43.42. Also my Athlon 1700+ (1450 Mhz) has been beating the PM Dual 866 G4 by quite a bit until i got my DVB-S card for the Athlon and started loosing processor time. I am running optimized clients on them two but the G4 and G5 should still do better and it looks like now they will.

Thanks you both again. You would have thought Apple would have done this a long time ago especially since they used it in demos at MacWorld with internally optimized clients. I am glad you guys stepped up to the task.

[Shaktai]

I'm running G5-A5 on my iMac G5 now, using the reference work unit. I'll also run G4-A5 just out of curiosity to see how much the optimizations differ. Hope I can remember how to read the time out of the results.

[electrician1972]

Have been running the G4 Alpha 5 clients since Sunday and have seen an average of 6500s on both of my 1.42 GHz Mac Mini's and about 8500s on my 1GHz iBook. EXCELLENT.... Chopped about 500s off each machine's run time.

No errors either. Thus far seems stable enough.

[Shaktai]

G5 iMac 1.6 ghz. 1 gig of RAM.

Okay, G5-a5 was 5678 seconds with the SETI Reference unit.
G4-a5 was 5717 seconds with the SETI reference unit.
Only 49 seconds variance.

[alexkan]

Originally Posted by Shaktai

G5 iMac 1.6 ghz. 1 gig of RAM.

Okay, G5-a5 was 5678 seconds with the SETI Reference unit.
G4-a5 was 5717 seconds with the SETI reference unit.
Only 49 seconds variance.

Out of curiosity, do you have timings for the reference unit with any of the older alphas?

[BTBlomberg]

Here are more extensive results from my use of Alpha-4:

Code:
PowerBook G4 500
Mac OS 10.4.2
768 MB Ram with lots of work activity as this is my main 
work machine as i need to move it around all the time.

Alpha-4 G4 Opt Seti 4.18 __________ Increase Multiple
Results Returned/Verified: _____ 9/6
Average CPU time (sec): ____ 9587.95 ____ 4.21
Average claimed credit: _______18.74
Average granted credit: _______19.17
Hours Per WU: _________________ 2.66 ____ 4.21
Credits Per Hour: _____________ 7.20 ____ 3.20
24 Hour Potential: __________ 172.73 ____ 3.20

Javalizard G4 Opt Seti 4.30 
Results Returned/Verified: ___ 16/16
Average CPU time (sec): ___ 40332.91
Average claimed credit: ______ 78.64
Average granted credit: ______ 25.23
Hours Per WU: ________________ 11.20
Credits Per Hour: _____________ 2.25
24 Hour Potential: ___________ 54.05


PowerMac G4 Dual 866
Mac OS 10.4.2
768 MB Ram and nothing but Safari, 
Dashboard and HP AIO Driver running.

Alpha-4 G4 Opt Seti 4.18 __________ Increase Multiple
Results Returned/Verified: ____ 15/8
Average CPU time (sec): ____ 8770.01 ____ 3.45
Average claimed credit: ______ 24.78
Average granted credit: ______ 24.50
Hours Per WU: _________________ 2.44 ____ 3.45
Credits Per Hour: ____________ 10.06 ____ 2.77
24 Hour Potential: __________ 241.34 ____ 2.77
Total 24 Hours for 2 CPU: ___ 482.69 ____ 2.77

Javalizard G4 Opt Seti 4.30
Results Returned/Verified: ___ 15/12
Average CPU time (sec): ___ 30248.17
Average claimed credit: ______ 85.48
Average granted credit: ______ 30.48
Hours Per WU: _________________ 8.40
Credits Per Hour: _____________ 3.63
24 Hour Potential: ___________ 87.07
Total 24 Hours for 2 CPU: ___ 174.14

I can't totally explain the Credit System of SETI but despite the increased crunching performance on WU between 345-421% the Credit System reflects only 75-80% of that increase. This makes me wonder if CPU time is factored into the credits or if this just happened to be a section of signal with less of what they factor into the Credits.

Also, I must wonder if the Alpha Client is more optimized for the type of G4 CPU (7410 v1.3) in the PowerBook opposed to the G4 CPUs (7455 v2.1) in the PowerMac as the Increase is less than I would expect between the two CPUs. The 866 should have 70% more to it I would have expected unless the OS Overhead is greater for the Dual CPU setup. Regardless the improvements are great.

[BTBlomberg]

I forgot to add I am using this with Boinc 4.44 Simple GUI Client on both.

[mikkyo]

Credit is granted based on other people's results and yours in a convoluted way that has little to do with your cpu speed or crunch time.
So mapping credits to the worker speed isn't worth much as a data point.
The speed is what matters, so assume faster worker gives you more WUs done, which gives you more credit over time.

[BTBlomberg]

Thanks Mikkyo,

Great explanation. I suspect that is why it takes longer for some to validate than others becasue of the distribution of the WU and getting all the results returned and then comparing them.

Again, it is good to see the estimate for Credit is more in line with reality in this version compared to the Javalizard version.

[alexkan]

Originally Posted by BTBlomberg

Also, I must wonder if the Alpha Client is more optimized for the type of G4 CPU (7410 v1.3) in the PowerBook opposed to the G4 CPUs (7455 v2.1) in the PowerMac as the Increase is less than I would expect between the two CPUs. The 866 should have 70% more to it I would have expected unless the OS Overhead is greater for the Dual CPU setup. Regardless the improvements are great.

Yes, I'm sure many of you have noticed that the performance of the alpha clients doesn't scale linearly with clock speed. This is because the SETI client is bound primarily by memory bandwidth/latency and L2 cache size on Macs, an unfortunate situation because the G4's front side bus is so slow. Those of you who have chips with 1 MB of L2 cache (mostly older G4s) should notice that those machines perform disproportionately well. This is because the largest FFT that the client performs, which is also the one it performs most often, has length 131072, which in this case requires 1 MB to hold the working set. Since the FFT walks all over that 1 MB, we'd ideally want it to be contained entirely in L2, since L2 is so much faster (especially in terms of latency) than going out to main memory. For everything else newer, we pay the price in speed for not being able to hold all our data in L2 at a time. This is a bit of a simplification because of virtual memory and the other processes running on the machine, but characterizes the problem pretty well.

Fortunately, Apple's taken care of this (at least as much as is possible) with the vDSP libraries, which is why we use their FFT code instead of the standard code that comes with the normal SETI distribution.

Of course, I can't help but wonder how G5s would perform were they to have 1 MB of L2 cache per processor instead of the 512 KB they actually have...

Also, upgrade to alpha-5 and see if it's even faster.

[Shaktai]

Originally Posted by alexkan

Out of curiosity, do you have timings for the reference unit with any of the older alphas?

No, but if they are still up on your website, I'll give one or two a try.

[BTBlomberg]

Alex,

Thanks for another great explanation and you are right, the PowerBook has a 1 MB L2 and the PM Dual G4 866 has a 256 L2 with a 1 MB L3.

Setting up Alpha-5 soon.

[Drash]

1GHz eMac, 10.4.2, 1GB RAM

g4-a5 seems slightly quicker than a4 roughly giving 8100-8300 ish with a single 7500, whereas a4 gave (over a larger sample) 8300-8500 with a smattering of 7600s

If you're interested http://setiweb.ssl.berkeley.edu/resu...hostid=1412422 sent 20 Sept and onwards is a5, 15-18 Sept is a4 and going backwards from 14 Sept is a3

Stilll going in the right direction, cheers ;-)