Scheduling Of Multiple Multi-Threaded Applications On CMP's

Summary:

Due to the limitations in the conventional processor designs, chip multiprocessors (CMPs), which have multiple cores on a single chip, are a promising alternative to singlecore architectures for performance improvements. The potential performance gains that can be achieved by the using CMPs decline when there is contention for the shared cache structure for multiple multi-threaded applications. Our main focus is to present mapping strategies of multiple multi-threaded applications on multicore architectures. We propose and develop a novel prediction-based mapping strategy. Our approach analyzes thread behavior of dierent applications on the shared cache by considering all possible thread combinations of dierent applications. It nds the best thread combinations of dierent applications that result in minimum cache disturbance. Our prediction-based framework has two components: a static component and a dynamic component. The collection of the training data which is given to the curve tting model as an input is done o-line at the static component. After receiving the predicted values, the threads of each application that shares the same core are arranged. Communication with curve tting model, receiving predicted results, and nally mapping according to these values are done on-line at the dynamic component. The communication between the application code and the curve tting model is provided by a runtime module which collects the training data from the application code and sends them to the curve tting model and receives predicted data from the curve tting model and sends them to the application code. Any interference with the program is avoided at every step of the execution.

Özet:

Klasik islemci tasarmndaki snrlamalardan dolay, tek bir yongada birden fazla cekirdege sahip olan Yongada Çoklu _Islemciler (CMP) performans gelisimi icin tek cekirdekli mimarilere umit verici bir alternatiftir. CMP kullanm ile elde edilebilir performans arts, coklu cokizlekli (multi-threaded) uygulamalarda paylasml onbellek yapsndaki cekismeden dolay azalabilir. Bizim esas odak noktamz, coklu cokizlekli uygulamalar icin haritalama stratejileri sunmaktr. Biz bu tezde, yeni bir tahmintabanl haritalama stratejisi sunuyor ve gelistiriyoruz. Bu yontem, farkl uygulamalarn izleklerinin paylasml onbellek uzerindeki davranslarn analiz eder, farkl uygulamalarn tüm izlek kombinasyonlarn tahmin eder, ve farkl uygulamalarn en az onbellek karsklgna sebep olacak en iyi izlek kombinasyonunu bulmaya calsr. Tahmin tabanl cercevemizin iki bileseni vardr: statik bilesen ve dinamik bilesen. Tahmin surecinin egitim asamas statik bilesende cevrimds olarak yaplr. Tahmin edilen degerler alndktan sonra, her uygulamadan kac adet izlegin ayn cekirdegi paylasabilecegi ayarlanr. _Ikinci bilesende ise, egri uydurma modeli ile iletisim kurulmas, tahmin sonuclarnn alnmas ve bu tahmin sonuclarna göre en son haritalama izlemi belirlenmesi calsma zamannda yaplr. Uygulama kodu ile egri uydurma modeli arasndaki iletisim runtime modulu tarafndan gerceklestirilir. Bu modul, egitim asamas icin gerekli olan bilgiyi uygulama kodundan alr, egri uydurma modeline iletir ve tahmin edilen bilgileri egri uydurma modelinden alr, uygulama koduna gönderir. Hicbir admnda programa karslmaz.