Tensorflow provides means to specify the number of CPU/GPUs to be used for each Session. Otherwise, it will try to use all CPUs available in the machine. Currently tensorframes does provide any means to specify this information and it ends up tensorflow using all the CPU cores in the executor nodes. This is more problematic when an executor has more than one cores assigned which results in more contention. An alternative would be to set only one core for each executor but this is not optimal in mix workloads where CNN inference is only a part of the workload.
Tensorflow provides means to specify the number of CPU/GPUs to be used for each Session. Otherwise, it will try to use all CPUs available in the machine. Currently tensorframes does provide any means to specify this information and it ends up tensorflow using all the CPU cores in the executor nodes. This is more problematic when an executor has more than one cores assigned which results in more contention. An alternative would be to set only one core for each executor but this is not optimal in mix workloads where CNN inference is only a part of the workload.