MilkyWay@home

[matob]

Vypustili niekoľko WU s novou app. V project preferences treba mat zaškrtnuté "run test applications"

[Hefto99]

Vydana je uz aj Linuxova verzia testovacej aplikacie:

Aktualny stav:
http://milkyway.cs.rpi.edu/milkyway/apps.php

[matob]

November 25, 2008
We've updated a bug or two in version 0.4, and also compiled with various optimization flags which should improve performance further. The code has been released for this version as well. We've also compiled a binary for 32 bit linux, so let us know how that works.

[matob]

November 30, 2008
I've released the version 0.5 code to the code release directory.

November 30, 2008
We're starting the move to only using the new application. No new workunits will be generated for the old application (however the old assimilator and validator will keep running until all the old workunits have been processed (so you'll still get credit for those). We'll be updating with a version 0.5 tonight which will fix the checkpoint error that's being output (it's just cosmetic). We'll also put out a binary for FreeBSD along with that. Since we've made the source code available, we'll only be supporting binaries for FreeBSD, Linux, OS/X and Windows, so if you're using something more obscure than that you'll have to compile it yourself because we don't have regular access to those types of systems. We've had some discussion and the new WUs should be a bit more computationally intense, which should help with the server problems we've been having. Not to mention that moving away from the old app and it's server side daemons will also help our server load issues.

--
-Vydaná bola nová verzia aplikácie - v 0.5 , aplikácia sa v boinc manageri zmení automaticky, stačí mať na projektovej stránke zaškrtnuté: run test applications (povoliť testovacie aplikácie).
-Od teraz bude podporovaná iba nová optimalizovaná aplikácia, pokiaľ sa nezrátajú všetky staré WU, za ktoré kredit dostanete.
-V aplikácii boli doladené kozmetické úpravy ako napr. checkpoint.
-Podporované systémy: FreeBSD, Linux, OS/X a Windows.
-WU budú výpočtovo náročnejšie a teda dlhšie, aby sa odľahčila záťaž na projektové servery.

[matob]

December 1, 2008
We've put the 0.6 windows binary on the server so it should be available now. Sorry for the wait. This should fix the checkpointing problems.

Kód bol aktualizovaný na verziu 0.6
-Oprava "padania" aplikácie pri štarte
-Oprava checkpointov

[matob]

Aplikácia v.0.7 prináša 10% zrýchlenie výpočtu.

[Hefto99]

Na Milyway prave prebieha hlasovanie o favicone (to je ta mala ikonka v prehliadaci vlavo od URL).

Kto chce moze sa zucastnit tu http://milkyway.cs.rpi.edu/milkyway/for ... php?id=546

Hlasovanie je neoficialne, je to aktivita cruncherov, ale myslim ze nebude problem presvedcit adminov projektu aby bola vybrata ikonka zaradena do projektovych stranok

[tahanko]

custom optimalizacia: http://milkyway.cs.rpi.edu/milkyway/for ... php?id=551 forum

odkaz funguje iba na SSE2 verziu, ostatne odkazy su mrtve.

optimalizacia funguje: pentiumM 1,7GHZ 5700s na 1600s (http://milkyway.cs.rpi.edu/milkyway/res ... ostid=9512)
a quad 9400: 2700s na 730s (http://milkyway.cs.rpi.edu/milkyway/res ... stid=39732)

[Honza]

SSE4.1 to by bylo konečně něco...a taky na vylepšení mé neutěšené kreditové bilance
Každopádně No work to send to jistí.

Takže zase hurá to práce, tentokráte na PG PPS3.

[tahanko]

SSE3 optimalizacia je stiahnutelna tu http://www.arizmoon.com. funguje velmi dobre

quad 9400: "old app" vs SSE2 vs SSE3: 2700s vs 730s vs 310s
http://milkyway.cs.rpi.edu/milkyway/res ... stid=39732

edit: na www.arizmoon.com najdete optimalizacie pre SSE, SSE2 a aj SSE3

[Honza]

Travis dnes napsal, co dělají nové analýzi.

You'll probably notice that there are workunits with new names available. All the different workunits are performing variations of the newton method ([html]http://en.wikipedia.org/wiki/Newton's_m ... timization[/html]) on the different stripes we're doing our tests on. In general what the newton method does is take the first and second derivatives of a function (in our case the likelihood function), and use those to find its minimum. This tends to converge to the minimum of the function VERY fast if it's close to the minimum, and in the case of a quadratic function it only takes one step (our likelihood function is not quadratic however).

What we're doing is taking all the points you're returning, and performing a regression analysis on them to estimate the first derivative (the gradient) and the second derivative (the hessian). If you remember any calculus, a point - gradient/hessian = root (if it's quadratic). In our case, we use point - gradient/hessian as a step to calculate the center of the next area we'll use for another regression. This process repeats until we can't optimize the function anymore.

Now, since our function is a lot more complex than a quadratic function, and since we're estimating the hessian/gradient through regression, the step calculated isn't necessarily the best we can do. This leads us to the different types of searches we're testing right now:

UR - The step taken is the same, however we update the range that parameters are generated in for the next regression on each step. The new range used is simply the new point +/- the step calculated. As we converge to an optimum, this method is rather effective (and simple), as the smaller the step, the smaller our range and more accurate our hessian/gradient calculations.

ER - This version also keeps the same step, but instead of updating the range based on the step, we calculate error values for each parameter in our function (based on the regression), and the next range used is 1-2 standard deviations of our calculations. We're interested to see if this lets us converge quicker and get better hessian/gradient estimates.

URLS - After calculating the hessian/gradient and using this to calculate the step for our search, instead of just taking the step, we use it as a direction to search along. In this case, we generate points along a line from the current point in the search along the calculated step. Using these we do another (more simpler one dimensional regression) to find the minimum along this line. This minimum is used as a more accurate version of the gradient/hessian step. Following this the range is updated as in UR.

ERLS - This performs a line search the same as URLS, except the range is updated as in ER.

[Kiwi]

MOjim snom bolo robit takeho algoritmy a ich optimalizacie, ale nie hrubou silou ale silou kreativity,
mozno sa raz k tomu dopracujem. Uvidime, ci uvidime.

[tahanko]

http://milkyway.cs.rpi.edu/milkyway/for ... d=551#8393 je podozrenie ze vysledky optimalizacie nie su presne, odporuca sa prejst na standardnu aplikaciu

[tahanko]

custom aplikacie na http://www.arizmoon.com/ su uz opravene, potvrdene je to aj od Travisa

[Honza]

Hlavně byla vydána oficiální 0.8, která konečně řeší memory leaks a mrvení výpočtů při checkpointu.
Nutno dodat, že nemalou díky těm, kteří optimalizují a opravují chyby v oficiální aplikaci.

EDIT: Bohužel, oficiální aplikace 0.8 pro 32-bit Linux nechodila vůbec, takže je stažená.