Design and Performance Evaluation of a Service-Oriented HLA RTI on the Grid - Grid Computing: Infrastructure, Service, and Applications - page 463

TABLE 18.1
Experimental Conigurations and Latency Results in ms
DMSO RTI
Test Federate1 Federate2 rtiexec + fedex Result

LAN c0-6 c0-7 c0-0 5.3

WAN pdcb1 Escience1 pdcb1 116.5

SOHR
Test

Federate1 LS1 Federate2 LS2 GT4.0.2 GT4.1.1

LAN c0-6 c0-4 c0-7 c0-5 85.6 19.6

WAN A Escience1 pdcb1 c0-6 c0-4 1026.8 241.6

WAN B c0-6 pdcb1 s30 Escience1 916.7 154.7

share the same TCP/IP connection for communication, so the overhead of
the connection setup and tearing down is amortized among the multiple
service invocations. Since SOHR's performance on GT4.1.1 is much better
than its performance on GT4.0.2, its performance results on GT4.1.1 are
used for comparison with the DMSO RTI in the following paragraphs.
On the LAN, SOHR's latency is 19.6 ms, which is 3.7 times that of the
DMSO RTI. SOHR's latency is larger than that of the DMSO RTI due to the
overheads of grid service invocations. Basically, grid service invocations
have two kinds of overheads, processing cost and communication cost. As
discussed in [22], the size of a SOAP message is around 1-2 kB per grid
service request/response, which is much larger than a DMSO RTI mes-
sage. This increases both processing and communication costs in SOHR.
To further analyze this, we carried out a test of a simple add(int a) service
invocation on the LAN using GT4.1.1, and found that service invocations
by clients executed on the same machine as the service and also on a dif-
ferent machine take almost the same time of 8 ms inside the cluster. This
means that the major overhead of the LAN experiment is the processing
cost while the communication cost is not high due to the fast intra-cluster
network connection.
Our SOHR was tested on the WAN using two coni gurations, with two
LSs deployed at the NTU side (WAN A) and one LS deployed at each side
(WAN B). WAN B has better performance, because the interval between
the time an interaction reaches an LRI (step 28 in Figure 18.4 ) and the time
the interaction is fetched by the receiving federate (step 30 in Figure 18.4)
depends on the service invocation sequence of the LIS. If the receiving
federate tries to fetch the interaction (step 30 in Figure 18.4) before the
interaction reaches the LRI (step 28 in Figure 18.4), this incurs extra time
and the effect is more prominent if a remote LS is used as in WAN A.
Sleeping time can be used by the receiving federate to delay the interaction