What Happens When Apple Passes Microsoft In Value? Yes, When.

vlada_vlada: Problem je sledeci: projekat zahteva da koristite slozene (npr. stablo-like) strukture podataka koje imaju veliku dinamiku izmene. Nakon odredjenog vremena elementi strukture ce biti rasuti duz velikog adresnog prostora (defragmentisani). Tada operacije nad strukturom, npr. iteriranje kroz elemente izazivaju dosta promasaja u keshu - sto zahteva dovlacenje blokova iz operativne memorije. I eto bottlenecka.

vlada_vlada: Drugi cest problem je sto se u C++-u programer stimulise da koristi dosta objekata.

vlada_vlada: Ove operacije pod windowsom verovali ili ne nisu thread safe. Win koristi globalni lock nad procesom da bi ih serijalizovao(!). Sto je tragicno za multithreaded aplikaciju.

In Apple's 2003 10-K filing with the SEC, the company revealed that Microsoft's $150 million investment bought the company 150,000 shares of Series A nonvoting convertible preferred stock at $1,000 per share. Microsoft had the option after August 5, 2000 to convert those preferred shares, for $8.25 per share, into common stock. In 2000, Microsoft converted a little under half its shares into 9 million shares of common stock. It then converted the remainder in 2001 into another 9.2 million shares.

All told, Microsoft spent a little over $151 million to acquire 18.2 million shares of Apple stock, for roughly $8.31 per share. Microsoft confirmed that it sold all of its AAPL holdings some time ago, and likely did so at a healthy profit—after all, AAPL has traded significantly higher than $8 for many years. But what if Microsoft had held on to that investment just a little longer?

As of approximately noon CDT Thursday, AAPL was trading at $250.99. That translates to a value of $4.57 billion, which would represent a staggering 3,000 percent ROI had Microsoft held on to those 18.2 million shares.

At its recent record peak at the end of April, AAPL was trading even higher, at $272 per share. Had Microsoft cashed out then, it would have had an extra $4.95 billion in its pocket—a nearly 3,200 percent ROI.

Since Steve Ballmer took over as Microsoft's CEO, the company's market cap has dropped to less than half its peak value, currently hovering around $219 billion. Ballmer recently downplayed the significance of the market cap drop by pointing out that Microsoft still earns far more profit that Apple. "I will make more profits and certainly there is no technology company in the planet which is as profitable as we are," Ballmer said earlier today during a press conference in New Delhi. "Stock markets will take care of the rest."

mmix: Problem sa heap globalnim spinlockom je resen u Windows 7/2008r2 kernelu. Drugo, to je ionako kacilo C++-evce ;) CLR (a i JRT) prealociraju memoriju u svoj intenrni heap koji vise ne zavisi od global locka, naravno uz space tradeoff.

Tacno, BP miss nas kosta onoliko koliko je dugacak sam instrukcijski pipeline, gubi se mnogo vise CPU ciklusa ako je pipeline duzi. Mada su oni isli na to da usavrse BP, AMD se sa K6 procesorima svojevremeno hvalio uspesnoscu pri BP-u od 95%.

Inlining moze tu dosta da pomogne, mada ne treba preterivati kao i prefetch kod vrednosti za koje znas da ce se nenormalno mnogo potrazivati u kodu. Ono sto mene nervira je da je GCC postao mnogo pametan i koristi heuristiku za inlining pa ti sam inline-uje ili uninline-uje inline-ovanu funkciju od strane programera i na kraju kad ti naleti neki bug ne znas da li je bug u toj liniji koda ili onoj 1K linija iznad gde je definisana inline funkcija. Muke sam imao sa kernelom sa GCC 4.5 i opcijom CONFIG_OPTIMIZE_INLINING, jedno vreme sam mislio da je GCC puko (sto i nije daleko od istine).

Cek zar i dalje, na linuxu se maksimalno iskoriscavaju spinlock-ovi, cak imamo i nekoliko varijanti:

http://www.mjmwired.net/kernel/Documentation/spinlocks.txt

Onda je i preempcija laksa ukoliko je locking mehanizam dobro implementiran unutar koda...

Ako projekat to zahteva, onda je to neminovno i ne zavisi od izbora programskoj jezika.

Ko te tera da koristiš više objekata nego što ti treba? U C++ se "(de)alokacija" uglavnom vrši na steku, dok ne potregneš new i delete ili strukture koje ih u sebi sadržr i kod kojih je to neizbežno. A neizbežno je iz malo fundamentalnijih razloga da bi to bilo specifično za programski jezik.

Mutex (un)lock je brza operacija ako se ne zahteva provera vlasništva niti nad mjuteksom, što OS neće ni da radi ako je mjuteks u njegovom vlasništvu. Problem može da nastupi samo ako dođe do zaglavljivanja jedne niti na njemu, jer ga je druga prethodno zaključala.

Ako se new i delete različitih niti u najvećem broju slučajeva promašuju, onda je to mnogo brža operacija od traženja slobodnog bloka i zauzimanja memorije.

Problem sa heap globalnim spinlockom je resen u Windows 7/2008r2 kernelu. Drugo, to je ionako kacilo C++-evce ;) CLR (a i JRT) prealociraju memoriju u svoj intenrni heap koji vise ne zavisi od global locka, naravno uz space tradeoff.

Pre ili kasnije trziste ce brutalno svesti stvari na pravu meru, masovna psihoza (hype) ce splasnuti i cene ce postati odraz realnih potreba - kao sto i treba da bude…

Sve u svemu - jos od Pentium M generacije i bacanjem NetBurst-a u zasluzeno djubre, Intel i te kako zna sta radi, i ima tehnologiju koja je prilicno ispred konkurencije.

vlada_vlada: Da se ponovim - bilo je neophodno upotrebiti dinamicke strukture. Naravno da je fundamentalno pogresno pominjati stek u tom kontekstu.

...samo kazem da su Power napredniji od intela - odnosno ono sto oni danas imaju, x86 ce imati za par godina. da li to nekome treba je vec druga stvar :)

Larrabee as an architecture is not dead however. Intel made it very clear that a variant of the architecture designed for High Performance Computing (HPC) applications was still on track. It’s called the Intel Many Integrated Core (MIC) architecture and it claims to leverage both Larrabee and Intel’s many core research projects. The first Intel MIC product (codenamed Knights Corner) will be built on Intel’s 22nm process and feature more than 50 cores. I’m guessing it’s too early to tell how many cores will yield once the chip is done and Intel doesn’t want to miss any targets so we get a conservative estimate today.

The 22nm requirement puts the earliest release for Knights Corner at 2011