c++ - Is it possible to execute multiple instances of a CUDA program on a multi-GPU machine? -

-

background:

i have written cuda program performs processing on sequence of symbols. program processes sequences of symbols in parallel stipulation sequences of same length. i'm sorting data groups each group consisting entirely of sequences of same length. program processes 1 group @ time.

question:

i running code on linux machine 4 gpus , utilize 4 gpus running 4 instances of program (1 per gpu). possible have program select gpu isn't in use cuda application run on? don't want hardcode cause problems down road when program run on different hardware greater or fewer number of gpus.

the environment variable cuda_visible_devices friend.

i assume have many terminals open have gpus. let's application called myexe

then in 1 terminal, do:

cuda_visible_devices="0" ./myexe

in next terminal:

cuda_visible_devices="1" ./myexe

and on.

then first instance run on first gpu enumerated cuda. second instance run on second gpu (only), , on.

assuming bash, , given terminal session, can make "permanent" exporting variable:

export cuda_visible_devices="2"

thereafter, cuda applications run in session observe third enumerated gpu (enumeration starts @ 0), , observe gpu as if device 0 in session.

this means don't have make changes application method, assuming app uses default gpu or gpu 0.

you can extend make multiple gpus available, example:

export cuda_visible_devices="2,4"

means gpus ordinarily enumerate 2 , 4 gpus "visible" in session , enumerate 0 , 1.

in opinion above approach easiest. selecting gpu "isn't in use" problematic because:

  1. we need definition of "in use"
  2. a gpu in use @ particular instant may not in use after that
  3. most important, gpu not "in use" become "in use" asynchronously, meaning exposed race conditions.

so best advice (imo) manage gpus explicitly. otherwise need form of job scheduler (outside scope of question, imo) able query unused gpus , "reserve" 1 before app tries so, in orderly fashion.


Comments

Post a Comment

Popular posts from this blog

PHP DOM loadHTML() method unusual warning -

-
i had following code works fine on localhost: $document = new domdocument(); $document->loadhtml($p_result); $form = $document->getelementsbytagname('form')->item(0); // code continues using $form variable after same code updated on outside server, loadhtml() failed , issued warning. warning: domdocument::loadhtml() expects parameter 1 valid path, string given in path/name/to/script.php returned null instead of object code pretty gets fatal error. note contents of $p_result absolutely same on outside server , on localhost. but why displays kind of warning , why not work? doesn't loadhtml() expects argument 1 string in first place? why method expects parameter 1 valid path ? just make clear i'm not calling loadhtmlfile() , i'm calling loadhtml() . thanks. you're affected one of php bugs . issue present in php 5.6.8 , 5.6.9. have affected php version on server, , bug-free version on localhost. the bug forbids null
Read more

python - How to create jsonb index using GIN on SQLAlchemy? -

-
here's current code creating index jsonb. index("mytable_data_idx_id_key", mytable.data['id'].astext, postgresql_using='gin') but got error. sqlalchemy.exc.programmingerror: (psycopg2.programmingerror) data type text has no default operator class access method "gin" hint: must specify operator class index or define default operator class data type. [sql: "create index event_data_idx_id_key on event using gin ((data ->> 'id'))"] is there way create index on sqlalchemy? the postgresql specific sqlalchemy docs @ http://docs.sqlalchemy.org/en/latest/dialects/postgresql.html#operator-classes mention postgresql_ops dictionary provide "operator class" used postgresql, , provide example illustrating use: index('my_index', my_table.c.id, my_table.c.data, postgresql_ops={ 'data': 'text_pattern_ops',
Read more

c# - TransactionScope not rolling back although no complete() is called -

-
i'm using transactionscope rollback transaction fail bool errorreported = false; action<importerrorlog> newerrorcallback = e => { errorreported = true; errorcallback(e); }; using (var transaction = new transactionscope()) { foreach (importtaskdefinition task in taskdefinition) { loader.load(streamfile, newerrorcallback, task.destinationtable, processingtaskid); } if (!errorreported) transaction.complete(); } i'm sure there no transactionscope started ahead or after code. i'm using entity framework insert in db. regardless state of errorreported transaction never rolled in case of error. what missing ? transactionscope sets transaction.current . that's does. wants transacted must @ property. i believe ef each time connection opened reason. probably, connection open when scope installed. open connection inside of scope or enlist manually. ef has nasty "design decision": default opens
Read more