@Fred
Switched over to Alex's Xeon v8 at 21 Apr 2008 4:06:45 UTC. Anything returned after that time stamp will have been crunched with the Xeon flavor.

BTW, I show 268 WUs with valid results (other than -9s) returned in the last 3.5 days. Two hundred of those WUs had angle ranges of 0.36 through 0.45.

QS

First results of the OSX Core2 vs Xeon Optimised Apps (and a slightly different format that I wanted to try for size):


Direct Link

Notes:
- The coloured columns are the number of WU's in each AR band; blue for Core2 and yellow for Xeon and relating to the scale on the left of the chart.
- For each column that is populated, the red line is the range of CPU times related to the scale on the right of the chart. The marker within the range indicates the mean value which is what I have been using in my previous posts.

From this it can be seen that the mean values from the relatively few Xeon results so far would seem to indicate better performance than Core2, but that all the results for Xeon are between the max and min times for Core2 in any given AR band. This demonstrates the caveat that I have so often quoted.

I would expect the bars for the Xeon to extend downwards as more readings become available over the next few days, pulling the mean value down to demonstrate better performance. But only time will tell.

Keep on crunching QSilver. We'll get you an answer (eventually).

F.

---

Notes:
- The coloured columns are the number of WU's in each AR band; blue for Core2 and yellow for Xeon and relating to the scale on the left of the chart.
- For each column that is populated, the red line is the range of CPU times related to the scale on the right of the chart. The marker within the range indicates the mean value which is what I have been using in my previous posts.

Thanks for the new format, There is quite a bit more information and it took me a couple glances to really see the benefit. I like it!

From this it can be seen that the mean values from the relatively few Xeon results so far would seem to indicate better performance than Core2, but that all the results for Xeon are between the max and min times for Core2 in any given AR band. This demonstrates the caveat that I have so often quoted.

What I see, with the limited data so far, is that the Xeon client CPU times are more consistant (within a tighter range). Possible explanation for this is lesser influence by other process running on the host.

I would expect the bars for the Xeon to extend downwards as more readings become available over the next few days, pulling the mean value down to demonstrate better performance. But only time will tell.

Maybe, but if the CPU times remain so tightly grouped and less than the Core2 mean, doesn't that also indicate better performance?

Keep on crunching QSilver. We'll get you an answer (eventually).

Touche...

---

I know it's not related to your App JD, but on the Mac Pro, generally units of the same AR take within 200 seconds of one another to complete. If you look at any group of results for the Mac Pro you will find that finish times are very similar to one another.

For Example -

816051829 Success Done 3,756.17 52.25 pending AR - 0.431999
816051828 Success Done 3,752.63 52.25 52.25
816051826 Success Done 3,755.65 52.25 52.25
816051824 Success Done 3,747.92 52.25 52.25
816051821 Success Done 3,757.21 52.25 52.25

Obviously this is just a very small sample but I think this is what you are trying to say.

---

Notes:
- The coloured columns are the number of WU's in each AR band; blue for Core2 and yellow for Xeon and relating to the scale on the left of the chart.
- For each column that is populated, the red line is the range of CPU times related to the scale on the right of the chart. The marker within the range indicates the mean value which is what I have been using in my previous posts.

Thanks for the new format, There is quite a bit more information and it took me a couple glances to really see the benefit. I like it!

From this it can be seen that the mean values from the relatively few Xeon results so far would seem to indicate better performance than Core2, but that all the results for Xeon are between the max and min times for Core2 in any given AR band. This demonstrates the caveat that I have so often quoted.

What I see, with the limited data so far, is that the Xeon client CPU times are more consistant (within a tighter range). Possible explanation for this is lesser influence by other process running on the host.

I would expect the bars for the Xeon to extend downwards as more readings become available over the next few days, pulling the mean value down to demonstrate better performance. But only time will tell.

Maybe, but if the CPU times remain so tightly grouped and less than the Core2 mean, doesn't that also indicate better performance?

Keep on crunching QSilver. We'll get you an answer (eventually).

Touche...

@John,
One of the reasons that I included the column to highlight the sample size. For example, in the 0.4 - 0.5 band there are 10 results for Xeon against 92 for Core2 which could contribute to the tighter grouping. If the grouping remains tighter as the Xeon sample size grows to match, and exceed, that for the Core2 then that would indicate, to me, better matching of the App with the architecture.

F.

---

I know it's not related to your App JD, but on the Mac Pro, generally units of the same AR take within 200 seconds of one another to complete. If you look at any group of results for the Mac Pro you will find that finish times are very similar to one another.

For Example -

816051829 Success Done 3,756.17 52.25 pending AR - 0.431999
816051828 Success Done 3,752.63 52.25 52.25
816051826 Success Done 3,755.65 52.25 52.25
816051824 Success Done 3,747.92 52.25 52.25
816051821 Success Done 3,757.21 52.25 52.25

Obviously this is just a very small sample but I think this is what you are trying to say.

Aw! SATAN you have spoiled my next surprise.

Look out for the same format for Mac vs Win.

F.

---

I know it's not related to your App JD, but on the Mac Pro, generally units of the same AR take within 200 seconds of one another to complete. If you look at any group of results for the Mac Pro you will find that finish times are very similar to one another.

For Example -

816051829 Success Done 3,756.17 52.25 pending AR - 0.431999
816051828 Success Done 3,752.63 52.25 52.25
816051826 Success Done 3,755.65 52.25 52.25
816051824 Success Done 3,747.92 52.25 52.25
816051821 Success Done 3,757.21 52.25 52.25

Obviously this is just a very small sample but I think this is what you are trying to say.

That is exactly the point I was trying to get across... In QS case, by switching to the Mac "xeon" client he is seeing a similar "narrowing" in the range of CPU times compared to when he was running the "core2" flavor in the same bins.

Of course this observation is made with only a limited # of results reporting. The ranges could spread as more variety of ARs within the bins are crunched.

A scatter plot, with points representing each result and vertical lines representing bin boundaries would bring more meaning to the range of CPU times per bin with these preliminary results. ( Just what our excellent chart people need to hear, more extraneous drivel from the peanut gallery ;-) [runs and hides before someone can toss an eraser at me]

Cheers,
JDWhale

---

That is exactly the point I was trying to get across... In QS case, by switching to the Mac "xeon" client he is seeing a similar "narrowing" in the range of CPU times compared to when he was running the "core2" flavor in the same bins.

Of course this observation is made with only a limited # of results reporting. The ranges could spread as more variety of ARs within the bins are crunched.

A scatter plot, with points representing each result and vertical lines representing bin boundaries would bring more meaning to the range of CPU times per bin with these preliminary results. ( Just what our excellent chart people need to hear, more extraneous drivel from the peanut gallery ;-) [runs and hides before someone can toss an eraser at me]

Cheers,
JDWhale

I've been trying to work out how to do something like that for several weeks now, without success so far. Standard Excel error bars are fixed percentages so do not help. My new chart uses the "Stock chart" intended for displaying variation in investment performance - with a few manual tweaks.

If anyone has any constructive ideas on this, please shout and we can start a new thread.

F.

---

I'll try and make one for my recent results, to see how it works out.

---

It seemed sensible to integrate my pictures into JDWhale's narrative on this occasion:

How about a weekly update with just minimal babble...

With the switch to RC1 SSE3 estimated RAC should be updated once again for Skipper, Lovey, and Wrongway; but I haven't plugged the data into my unofficial guestimator yet. We should be seeing some updated plots/charts if/when the data fills in the AR bins.

Host = Skipper: P4D-820 @ 2800MHz


06Apr My estimated RAC = 1280
13Apr My revised estimated RAC = 1342

Direct Link

The pre-release Chicken continues to perform for SSE3. Looking good. I will "lose" Whaleport column in future as it has served its purpose.

Host = Lovey: E4500 @ 2420MHz


06Apr My estimated RAC = 2300
13Apr My revised estimated RAC = 2720

Direct Link

The E4500 seems to prefer "Xeon" build and benefits from SSSE3 compared to SSE3.

Host = Wrongway: Q6600 @ 2520MHz


06Apr My estimated RAC = 4700
13Apr My revised estimated RAC = 5040

Note here... As I'm writing, Wrongway is still running benchmarks for the Lunatics crew. After review of said benchmarks, I'll switch BOINC to run an RC1 client of my choosing (probably the one with the best performance in the mid-AR WUs we're seeing so many of.

Warning : if there is an onslaught of VHAR or VLAR WU's I won't hesitate to switch to the best client for the data. Same holds true for Lovey.

Direct Link

Again, the Q6600 benefits from the "Xeon" build. Not sure if the "Xeon" switches were flicked in the compiler for the build running here. Note that the one AR band where the Chicken App appears to be slower than the Whale Xeon App represents just 3 (possibly unrepresentative) results from the Chicken App.

Host = Thurston: Q6600 @ 3375MHz


06Apr My estimated RAC = ~7400
13Apr My revised estimated RAC = 7080

Remember - Random unattended "Reset project" ~40 hours ago :-(
At least it's going in the right direction!

However tempting to switch... Thurston will remain on AK_WhalePort V0.2X until official Lunatics release... I want to see how accurate my estimated RAC of 7080 is.

Direct Link

Again we can see the benefit of the "Xeon" build for the Q6600. Interesting to note how little effect upping the speed to 3375MHz had.

Host = Ginger: P4-540 (Prescott H/T) @ 3200MHz

Sorry, Ginger is in the powder room and doesn't want her picture took.

Trust me, her graph is climbing, but since she's been switching between KWSN R-2.4 (SSE3) and AK_WhalePort V0.2S (SSE3), and now running RC1 (SSE3), it's not a reflection of a single clients performance. Trust the charts/graphs posted by the experts.

Direct Link

Ginger seems to be thriving on her new Chicken diet...

F.

---

I've been tracking the results on my own as well. Through 8:30AM CDT (1330 UTC), 58 WUs have been crunched & returned using Alex's v8 Xeon. That excludes 2 units returned with -9s. I had 267 WUs since I started tracking Core2 results last week. FYI--other usage on this MBP has been fairly consistent between the Core2 and Xeon crunch periods.

I've been focusing on the 0.36-0.45 AR because that's where most of the results are (Core2 had 203, Xeon was up to 49 WUs this morning). The Xeon average is just slightly worse than the Core2's: 4380 vs 4363. Xeon is about 5% faster in the lower end of that range and about 2% worse at the higher end of that range.

I looked closely at the 0.387-0.396 AR because there's a similar number of results for both v8s (Core2 had 33, Xeon has 37 so far). There's much less variation in times: Core2's range was 4367 to 4947 with an average of 4769 vs. Xeon's range of 4502 to 4620 with an average of 4534.

I can post my scatter graphs if someone can point out instructions to ensure small(er) pictures are displayed within this thread.

QS

---

Thurston is positioned at #21 on the "Top Hosts" list as I write this.

With some help from our fallen friend, #20 appears guaranteed within the next day or so.

There is one host I'm in fear of jumping in ahead of me, UL1' Big Mac. Though currently sitting in #31, this host has the ability to leap several spots when fed the proper diet.

I am begging The Curmudgeon to throw down the gauntlet if need be. Do not let Big Mac past.

The quest continues,
JDWhale

---

Lol,

Well over the next most of you will pass my mac, I'm going away for a week and turning the machine off apart from a few hours every morning.

Will still be online with a laptop but very few results will be posted.

---

Just the facts!

This reporter has just been informed of The Definitive Answer to the longstanding question of :

Which client is faster on the same hardware: OSX or WinXP ?

The pretty pictures will follow, but I'll give a brief account of what I have been made privy to...

It is too close to call!!!

It is a "photo finish" folks.

DO NOT TEAR UP YOUR BETTING SLIPS!!!

We are waiting for the photos to be developed.


Personal thanks to :

• UL1 for providing access to his Mac Pro 8-core X5365 ('lil Mac) for testing.
  
• Fred W for monitoring 'lil Mac and providing analysis.
  
• Jason G for allowing us to run the R28 beta clients "live".
  

for making this test possible.

A few more Mac Pro owners were invited to participate, but for one reason or another were unable to.
Thanks SATAN and Adrian for your consideration.

Pretty pictures to follow after our professional catches some ZZZZs.
(It's way past Fred' bedtime already.)

Regards,
JDWhale


[Edit] I see that in Fred' absence, I've been made proxy :

.
.


There you have it... Blaming the poor distribution of ARs in the distributed WUs for the testing period we can only see that the Windows client has a slight edge over the ARs in bins either side at 0.4, while OSX holds a slight advantage at VLAR.

It is this reporters opinion that The Playing Field is Now Level.

Let the fun begin.

[/edit]

JD;
I've a dual Quad core Mac under OS 10.5.2 that I can offer to do some testing with for you if it is still need/wanted.

Russell

JD;
I've a dual Quad core Mac under OS 10.5.2 that I can offer to do some testing with for you if it is still need/wanted.

Russell

Thanks for the offer... First a question back at you... Is your signup date really earlier than mine? I find it hard to trust anyone claiming to have joined the project before me.

Just kidding....

With the pending release the new Windows clients only about a week away, It would be hard to gather a meaningful sample in the limited time. The folks monitoring hosts already have more on their plate than they can comfortably chew. Add to that the requirement of your being able to switch between OSX and WindowsXP/Vista on your MacPro as/when directed make this a difficult proposition.

I hope you understand,
JDWhale

[Edit] Wait just a minute... I just spotted you at #30 spot and rising really, really fast. Does your Mac do Windows? Maybe I can sidetrack you to allow me to get even further up the ladder ;-)
[/edit]

---

JD;
I've a dual Quad core Mac under OS 10.5.2 that I can offer to do some testing with for you if it is still need/wanted.

Russell

Thanks for the offer... First a question back at you... Is your signup date really earlier than mine? I find it hard to trust anyone claiming to have joined the project before me.

Just kidding....

With the pending release the new Windows clients only about a week away, It would be hard to gather a meaningful sample in the limited time. The folks monitoring hosts already have more on their plate than they can comfortably chew. Add to that the requirement of your being able to switch between OSX and WindowsXP/Vista on your MacPro as/when directed make this a difficult proposition.

I hope you understand,
JDWhale

[Edit] Wait just a minute... I just spotted you at #30 spot and rising really, really fast. Does your Mac do Windows? Maybe I can sidetrack you to allow me to get even further up the ladder ;-)
[/edit]

Sidetrack him? What's a matter You want to play, Ketchup? Heinz old chum? chomp, chomp. ;) I think You're talking about rac here and not actual points, Of course the Dark Side of Seti calls or rather a movie does(I needed to test the TVs sound, Sat box needed a reset as It went mute upon Me starting the Sat box, weird, For a second I thought either the TV or a cable or something went dead, It's all I'd need as now I have to get an estimate on the 1st for repairing a scrape(It went to the plastic only) on the lower right front bumper cover of the Mustang GT and the place is about 14 miles one way from here(Gas$$)). :(

---

The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's

oupps ... my 2nd machine is running the merge of 2 versions of my code+ 2.4, I forgot to adjust the multiplier, please don t get mad, i just forgot.Will be fix in morning.


V

oupps ... my 2nd machine is running the merge of 2 versions of my code+ 2.4, I forgot to adjust the multiplier, please don t get mad, i just forgot.Will be fix in morning.


V

[Edit] Post # 777 Wish I was playing the slots... [/edit]

Maybe you refer to result like wuid=249775937.
I understand, simple mistake. You only get "extra" credit if your wingman also "overclaim".

Thank you for sharing...

Your SKT are looking very good and climbing very fast.

Congratulations to you!

Regards,
JDWhale

---

Well, after a short run on SSE4.1, the Frozen Penny was switched to the SSSE3 pre-release build. Results to date are as follows:


Direct Link.

165 of the 182 WU's collected for the SSSE3 App to date are in the AR bands immediately adjacent to 0.4 (and that is where 243 from a total of 248 collected for the SSE4.1 App were). Although the difference in average CPU times is discernable from the chart, the difference in max CPU times between the Apps in these two AR bands is less than half of one percent. They are as near identical as I would expect ever to get without re-crunching the same WU's (and a difference of this magnitude could even occur then).

F.

---

Just to complete the picture from the test-pilot-in-chief, here are the latest results from the Frozen Penny's little brother (Q6600):


Direct Link

Most of the results from the latest build are around the 0.4 AR with a significant number appearing at VLAR.

Still looking good :)))

F.

---

bump

With the switch to RC1 SSE3 estimated RAC should be updated once again for Skipper, Lovey, and Wrongway; but I haven't plugged the data into my unofficial guestimator yet.

Sorry still no updated estimates for RAC on these hosts. I've been busy trying get the holes on the data samples filled on Lovey to paint a more complete picture by our wonderful statisticians. With my running Lunatics RC1 on these hosts I agree that the value of AK_WhalePort displays are becoming of less interest. With the data painting a nearly complete fill(given the availability of variation in WUs), monitoring of Skipper, Ginger, Lovey and Wrongway should be terminated at your discretion. FYI : I will pull the plug on Ginger when her cache empties in a couple days.

My focus is now pointed at "burning in" the release candidates so there are fewer (hopefully none), surprises upon release.

A brief update on 2 of my hosts:

Host = Wrongway: Q6600 @ 2520MHz


06Apr My estimated RAC = 4700
13Apr My revised estimated RAC = 5040
23Apr Newest revised estimate RAC = 5400
I realize that Wrongway is now running RC1 SSSE3X flavor, this fact has minimal (if any) impact on my evaluation. My earlier estimates on this host were intentionally conservative. This RAC 5400 number is the number you should expect if you are running a "stock clocked" Q6600 with DDR2-800 dual channel memory and the new client.


Host = Thurston: Q6600 @ 3375MHz


06Apr My estimated RAC = ~7400
13Apr My revised estimated RAC = 7080

Recovery from the "Random Project Reset" 3 days past was fairly smooth, helped with the immediate DL of 80 reissued WUs that I used to my advantage to minimize "pending credit" build. Please understand... I am not using BOINC default schedueling with Thurston. I am employing techniques to minimize pending credits. I have only cancelled a couple WUs that were already validated by late wingmen and have refrained from cancelling WUs where my wingam was running BOINC 4.45. My employed methods are intended to reach "final stable" RAC in as little time as possible and are not intended to inflate RAC beyond what would be attained with more available time. YMMV

Estimated RAC 7090 remains for Thurston. This should be your expected performance of a Q6600 @ ~3300MHz with DDR2-1066 dual channel memory.

FYI: On the jukebox Frank by "Amy Winehouse"... How can anybody resist her sultry voice and honesty? IMO, totally deserving of her Grammy as "Best New Artist".

Still searching for the "Nuclear Wessels",
JDWhale

---