Adaptive polling
A distributed computing system manages execution of jobs and their associated tasks. A broker manages assignment of computing tasks from clients to available computing resources. Clients and available computing resources contact the broker by polling. To prevent “ringing,” the broker specifies wait times for the polling entities, and randomizes the wait times in a range around a desired target latency. That is, a pseudo-random number generator is used to select values within a range of the target value, to avoid the situation in which deterministic patterns in the polling and response times result in highly synchronized message traffic, which might otherwise overwhelm the broker and/or the communication network.
1. A method for operating a task broker in a network-based distributed computing system, the method comprising:
receiving, from a processing resource, an indication of availability and, in response thereto;
indicating, to the processing resource, a first pseudo-random wait time, after which the computational resource may again indicate its availability to the broker;
receiving, from a job-submitting resource, an inquiry concerning the completion status of a job and, in response thereto indicating, to the job-submitting resource, that the job is not yet complete and indicating a second pseudo-random wait time, after which the job-submitting resource may again inquire about the completion status of the job; and
selecting at least one of the first or second random wait times using a pseudo-random number generator, wherein a target distribution of the pseudo-random number generator is adjusted in response to the number of processing resources and job-submitting resources in active communication with the broker;
wherein the pseudo-random number generator varies its target distribution in response to detected levels of network traffic in the distributed computing system;
wherein the pseudo-random number generator generates the target distribution of the first pseudo-random wait time or the second random wait time using an algorithm to vary the first pseudo-random wait time or the second pseudo-random wait times so as to maintain a total aggregate polling rate within an operating range such that when more customer entities are active, the pseudo-random number generator increases the first pseudo-random wait time or the second pseudo-random wait time, and when fewer customer entities are active, the pseudo-random number generator decreases the first pseudo-random wait time or the second random wait time.
2. The method for operating a task broker, as defined in claim 1 , further comprising selecting the first and second random wait times using the pseudo-random number generator.