1
1

11 Коммитов

Автор SHA1 Сообщение Дата
Edgar Gabriel
92d1b99468 optimize the shuffle step:
1. use communicator collectives if possible for performance reasons
 2. combined multiple allgathers into a single one
2016-02-19 11:04:04 -06:00
Edgar Gabriel
e63836c653 clean up the mca parameter handling of the component. Add new parameters for number of sub groups and write chunk size. This will allow to perform a systematic parameter study. 2016-02-19 10:15:28 -06:00
Edgar Gabriel
268d525053 change the tag to be a positive value. handle 0-byte situations correctly. 2016-02-19 08:28:50 -06:00
Edgar Gabriel
ad79012059 first cut on the version which overlaps the communication/computation of 2 iterations. 2016-02-19 08:28:50 -06:00
Edgar Gabriel
b253d4e887 fix CID 1349739, CID 1349738, CID 1349736 and (probably) CID 1349740 (not entirely sure about the last one, since I don't understand why block[i] is a problem but max_len[i] allocated and treated exactly the same way 1 line later is not). 2016-01-21 08:32:23 -06:00
Edgar Gabriel
a9ca37059a improve the communicaton abstraction. This commit also allows all aggregators to work simultaniously, instead of the slightly staggered way of the previous version. 2016-01-17 09:48:49 -06:00
Edgar Gabriel
26c57ef374 separate the size of the buffer used for the shuffle step and the size of the buffer used for a pwritev operation. 2016-01-17 09:48:49 -06:00
Edgar Gabriel
39d5c8c281 further bug fixes silencing a compiler warning and fixing a memory overrun 2016-01-17 09:48:49 -06:00
Edgar Gabriel
2bcae84e11 further debugging 2016-01-17 09:48:49 -06:00
Edgar Gabriel
2bdd6ba17a correctly free some buffers, and ensure that lustre_stripe_size and stripe_count are always read from the file system. 2016-01-17 09:48:49 -06:00
Edgar Gabriel
d282e94b67 add the new dynamic_gen2 component, designed to coexist for now with the original dynamic component 2016-01-17 09:48:49 -06:00