Sun T2000 används av forskare

2006-05-17 Peter Gustafsson   Lyckade lösningar, Universitet & högskolor

Som ni kanske vet försöker vi stötta univeritet och högskolor på alla möjliga sätt. Mest handlar det om speciella erbjudanden och kampanjer (finns alltid på http://www.inserve.se/inserve/edu) men ibland händer det spännande saker på forskningssidan.

Nu senast handlar det om T2000 (Suns kodnamn Niagara) som används i ett exjobb på Chalmers: “An Automatic Parallelization Tool Based on Thread-Level Speculation” (Alexander Busck och Mikael Engbom).

Forskare i Uppsala har också kunnat verifiera att man delvis kommer att behöva tänka om när man tar fram parallella algoritmer i framtiden. Ett resultat av detta arbete är att man lyckats få med en presentation på den prestigefyllda konferensen International Conference on Supercomputing, ICS06, i sommar. Nedan en kort intro för den som är intresserad: “Multigrid and Gauss-Seidel Smoothers Revisited: Parallelization on Chip Multiprocessors” (Dan Wallin, Henrik Löf, Sverker Holmgren, Erik Hagersten, Uppsala university, Sweden)

Efficient solutions require a match between the algorithm and the underlying architecture. The new chip-multiprocessors, CMPs (a.k.a. multicore), feature low intra-chip communication cost and smaller per-thread caches compared to earlier systems. From an algorithmic point of view this means that data locality issues become more important than communication overheads. This may require re-evaluation of many existing algorithms.

We have investigated parallel implementations of multigrid methods using a new temporally blocked, naturally ordered, smoother implementation. Compared with the standard multigrid solution based on the two-color red-black algorithm, we improve the data locality often as much as ten times while our use of a fine-grained locking scheme keeps the parallel efficiency high.

While our algorithm initially was inspired by CMPs, it was surprising to see our OpenMP multigrid implementation run up to 40 percent faster than the standard red-black algorithm on an 8-way SMP system. Studying the smoother part of the algorithm in isolation often shows it performing two iterations at the same time as a single iteration with an ordinary red-black smoother. Running our smoother on a 32-thread UltraSPARC T1 (Niagara) CMP demonstrates the communication cost of our algorithm to be low for such architectures.

Technorati tags: