ray.wait doesn't prioritize returning tasks that have finished earlier

[ericl]

This might be intended behavior, but naive use of ray.wait() can starve workers from having their results processed. For example, in this simple example where tasks take basically no time and the driver processes them slowly one at a time, some results can get delayed for hundreds of wait steps.

import ray
import sys
import time

@ray.remote
def f(i):
    return i

ray.init(num_cpus=4)

i = 0
remaining_ids = [f.remote(i) for _ in range(5)]

for _ in range(1000):
    [ready], remaining_ids = ray.wait(remaining_ids, 1)
    i_old = ray.get(ready)
    print(i - i_old, end=" ")
    sys.stdout.flush()
    time.sleep(.01)  # do some work e.g. apply gradients
    i += 1
    remaining_ids.append(f.remote(i))

Output:

0 1 2 1 4 4 1 4 1 9 1 1 1 1 1 1 1 1 1 11 1 1 16 1 3 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 21 1 1 1 1 1 1 1 48 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 49 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 74 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 223 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1 1 75 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 

Note that you can work around this by waiting on only the first element of the list.

[virtualluke]

This might be what I was seeing in #948

[robertnishihara]

Right, because ray.wait doesn't prioritize tasks that finished earlier (that's what you mean by fairness, right?), I think it actually just prioritizes objects that are earlier in the list of object IDs passed in (is that what you're seeing?).

Do you think this is a problem?

[ericl]

It's kind of odd, and basically means ray.wait() is unusable in stale-asynchronous models of communication without more effort (you will need to carefully manage which futures to pass to ray.wait to bound staleness).

We should at least document it; note also that Java concurrency primitives typically provide both a fair and unfair mode (unfair can be better for performance): https://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ArrayBlockingQueue.html

[ericl]

Btw, it's not the case that objects earlier in the list are prioritized -- in the above example, I always append to the end of the list, but still see high variance in delays.

[ericl]

I'm seeing this issue also in ray.tune, where if the driver pauses between wait calls to print out expensive debugging info, the queue buildup causes starvation. Here the 26x26 task has obtained nearly 100x more time than the other tasks:

 - PPO_FrozenLakeDeterministic-5x5-v0:  RUNNING [pid=29027], 30 s, 12039 ts, 0.618 rew
 - PPO_FrozenLakeDeterministic-6x6-v0:  RUNNING
 - PPO_FrozenLakeDeterministic-8x8-v0:  RUNNING [pid=30237], 18 s, 4063 ts, -0.592 rew
 - PPO_FrozenLakeDeterministic-11x11-v0:        RUNNING [pid=30842], 170 s, 96727 ts, 8.2 rew
 - PPO_FrozenLakeDeterministic-15x15-v0:        RUNNING
 - PPO_FrozenLakeDeterministic-20x20-v0:        RUNNING
 - PPO_FrozenLakeDeterministic-26x26-v0:        RUNNING [pid=32555], 1377 s, 28202 ts, -0.524 rew

(the workaround here is to not print out the debug info, but at larger scales you could see queue buildups even without this sort of time).

[luochao1024]

@ericl Does this issue only occurs when task is basically doing nothing but return i? If I let each task sleep(0.1), It seems that ray.wait() works properly. Here is the result:

import ray
import sys
import time

@ray.remote
def f(i):
    time.sleep(0.1)
    return i

ray.init(num_cpus=4)

i = 0
remaining_ids = [f.remote(i) for _ in range(5)]

for _ in range(1000):
    [ready], remaining_ids = ray.wait(remaining_ids, 1)
    i_old = ray.get(ready)
    print(i - i_old, end=" ")
    sys.stdout.flush()
    time.sleep(.01)  # do some work e.g. apply gradients
    i += 1
    remaining_ids.append(f.remote(i))

output
0 1 2 3 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4

[ericl]

I think that's because the main loop can keep up with the tasks in that case. If there is a backlog of tasks to fetch then the problem arises.

…

On Tue, Jun 19, 2018, 10:23 AM luochao1024 ***@***.***> wrote: @ericl <https://github.com/ericl> Does this issue only occurs when task is basically doing nothing but return i? If I let each task sleep(0.1), It seems that ray.wait() works properly. Here is the result: import ray import sys import time @ray.remote def f(i): time.sleep(0.1) return i ray.init(num_cpus=4) i = 0 remaining_ids = [f.remote(i) for _ in range(5)] for _ in range(1000): [ready], remaining_ids = ray.wait(remaining_ids, 1) i_old = ray.get(ready) print(i - i_old, end=" ") sys.stdout.flush() time.sleep(.01) # do some work e.g. apply gradients i += 1 remaining_ids.append(f.remote(i)) output 0 1 2 3 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1128 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAA6SmitSTGMMOZs0eL1C02NQWWhIM--ks5t-TOqgaJpZM4P5kwO> .