disk read: If the requested file is not cached in any node, the file needs to be read
from the disk. The initial node wakes a helper process, which reads the requested
The PRESS model has the high cache hit ratio because it uses the whole accu
mulated cache of a cluster. But the disadvantage of this is that the remote cache read
requests have long round trip time compared to the the local cache read request. Con
sidering that high percentage of requests have to be served as the remote cache read re
quests, this longer round trip time can give the clients dissatisfaction for the Web service.
Therefore, our goal is to reduce the round trip time for the remote cache read requests
without any penalty in the web server. We achieve this goal using the coscheduling
schemes in the Web server. Figure 3.2 shows the architecture of the proposed cosched
uled PRESS model. Like the PRESS model in Figure 3.1, the model has two processes or
threads involved in intra cluster communication. We do not distinguish processes from
threads in our simulation, since in the Linux operating system, one thread maps to one
process at the kernel level [25, 17]. These two intra cluster communicating processes are
blocked when there is no request to send or receive.
In this model, a remote request is forwarded by the remote cache send process.
The main process puts a request in a queue. Since we use a coscheduling mechanism
to service the remote cache read requests, the remote cache send process handles one
remote cache request at a time depending on the underlying coscheduling algorithm.
Although we have simulated most of the prior coscheduling algorithms, in this thesis,
we discuss only two such algorithms: dynamic coscheduling (DCS)  and DCS with
blocking. The DCS scheme works under the premise that when a request arrives at a