Well, actually is an HP RX7640 We don´t use node memory availability to balance job placement right now. We are studying the influence of thread/core/memory placement on the performance of different codes and the results are quite interesting (depending on the type of test, it´s more efficient to put the threads on the same "node" or in separates "nodes". In some cases the default kernel placement of threads in the system can be improved). We plan to implement this mechanism with cpusets (or maybe with taskset)
Carlos Fernandez Sanchez, PhD Systems Manager Centro de Supercomputacion de Galicia (CESGA) Avda. de Vigo s/n. Campus Sur 15705 Santiago de Compostela SPAIN Tel.: (+34) 981569810, ext. 232 P Before printing, think about ENVIRONMENTAL responsibility! -----Mensaje original----- De: Mark Hahn [mailto:[email protected]] Enviado el: lunes, 26 de enero de 2009 17:02 Para: Carlos Fernandez Sanchez CC: 'Beowulf Mailing List' Asunto: RE: [Beowulf] ECC exerciser/exorciser? > Regarding numactl, we are using it and shows good results: > > numactl --hardware > available: 3 nodes (0-2) > node 0 size: 57280 MB > node 0 free: 32579 MB > node 1 size: 57255 MB > node 1 free: 55920 MB > node 2 size: 16371 MB > node 2 free: 16140 MB > > (on a 128GB system, SLES10SP1) sounds like an altix. yes, it works fine on our 64 node altix as well. what I'm wondering is how widespread it is to find commodity PCs that have inconsistent numactl output. incidentally, do you use node memory availability to balance job/thread placement? _______________________________________________ Beowulf mailing list, [email protected] To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
