Decentralized input/output resource management

A tangible computer-readable medium embodying program instructions for managing a shared IO resource that is accessed by a plurality of hosts that includes a first host and a second host, the program instructions causing execution of a method, the method comprising:

• at the first host, monitoring latencies of IO requests issued by the first host to the shared IO resource, computing a first host latency value that is normalized with respect to an average IO size of the IO requests issued by the first host, and storing the first host latency value, wherein the first host is assigned a first number of shares to the shared IO resource in accordance with a quality of service (QoS) policy;
• at the second host, monitoring latencies of IO requests issued by the second host to the shared IO resource, and computing a second host latency value that is normalized with respect to an average IO size of the IO requests issued by the second host, and storing the second host latency value, wherein the second host is assigned a second number of shares to the shared IO resource in accordance with the QoS policy;
• and
• executing a management process for the shared IO resource using the first host latency value and the second host latency value, including:
• adjusting a window size of the first host based on the first host latency value and the first number of shares, wherein the window size of the first host relates to the size of the IO request queue for the first host;
• and
• adjusting a window size of the second host based on the second host latency value and the second number of shares, wherein the window size of the second host relates to the size of the IO request queue for the second host.

The tangible computer-readable medium of claim 10, wherein the window size is adjusted according to the following formula:

• w
• 
• (
• t
• +
• 1
• )
• =
• (
• 1
• -
• γ
• )
• 
• w
• 
• (
• t
• )
• +
• γ
• 
• (
• LAT
• threshold
• NL
• 
• (
• t
• )
• 
• w
• 
• (
• t
• )
• +
• β
• )
• ,
• where w(t+1) is the adjusted window size for time t+1;
• w(t) is the current window size;
• γ is a constant value selected between zero and one and determines the influence that the current window size has on the new window size;
• LATthreshold is a system-wide latency threshold selected to balance throughput with latency;
• NL(t) is the computed latency value at time t;
• and β is a per-host value based on an assigned share representing a relative level of priority of the host relative to other hosts.