26

pending messages, which need to be sent or received > T

1

, then the remote cache send

process yields the CPU to other processes instead of waiting for a reply from a service

node. At a service node, the DCS coscheduling algorithm is applied only when the

number of open connections < T

2

(The number of open connections is widely used to

estimate the Web server load [58]). If so, the remote cache recv process is assigned to

the CPU according to the Linux scheduler policy only. The DCS coscheduling algorithm

is beneficial only when a sending process receives a reply within the waiting time. It

not, the CPU time is wasted. Therefore, the adaptive model deploys the coscheduling

mechanism only when the probability of receiving a reply within a time quantum is higher

than the threshold value. Note that the adaptive model does not have to exchange

additional information among servers since they use their local information. T

1

, the

number of pending messages, is obtained to count the VIA sending and receiving queues

and T

2

is easily calculated in the Web server programs. Here, we choose T

1

based on

the experiment results and T

2

by the CPU utilization. However, finding the optimal

threshold values is another future research topic.