google bigquery - QueueDequeue step is a significant bottleneck in Tensorflow code -
i've seen few other questions regarding dequeueing step causing bottleneck have tried many of suggestions without success:
- i using multiple threads based on cpu count
- i have tried small batches (of 100) , larger batches (of 1000)
- i have tried switching shuffle_batch_join , batch_join
none of these things seem helping overall time. extent dequeue step causing bottleneck in code seems worse others have experienced. when @ timeline of other steps practically disappear in comparison. wondering if partially using bigquery , bigqueryreader source although other people experiencing slowdown assuming isn't cause.
i'm not entirely sure how interpret chart doesn't seem the problem being caused empty queue.
additional info
- capacity set batch size * 10
- min after dequeue set batch size * 2 + 1
- enqueue many set true
does have thoughts on else might able try speed things while still using bigquery source of data?
Comments
Post a Comment