Information Technology Reference
In-Depth Information
π-Channel Throughput on a Multi-
Cluster Testbed
direct streaming. The performance results show
a reduced dependence on the π-Server to store
pipe segments.
Figure 7 shows the segment send rates while
Figure 8 presents the measured bandwidth. The
π-Server executed on edda, one application on
tango and another on wexstan. The best perfor-
mance is achieved under asynchronous operation
(scenario three). The detailed results are presented
on Table 7. The tests were conducted with at least
12 trials per case. Of these, the mean is computed
using ten results, discarding the highest and
lowest values. The standard deviation given is
for the throughput, i.e., the mean message send
rates for each of the segment sizes. Note the
absence of observable performance differences
for Scenario 1 and 2. This means that matching
pi_attach() with pi_create() operations does not
lead to any improvement in data transfer rates.
Thus, π-Space/π-channels applications may be
written without using an odd-even rule to match
reads and writes.
In Scenario 3, the use of pi_attach() notifies
the π-Server of a pending request to retrieve a
We evaluate and compare the data transfer through-
put using π-channels under three scenarios: (a)
matched create/attach operations; (b) create first,
then attach; and (c) asynchronous read opera-
tions. Two processes are executed, using a pair
of π-channels for communication.
For scenario one, processes use paired pi_at-
tach() and pi_create() operations, i.e., when one
process is writing, the other is reading. For
the second scenario, each process executes all
π-channel writes first, followed by reads. The
π-Server caches most of the pipe segments during
the write phase. Once the processes perform the
pi_attach(), the pipe segments are retrieved from
the π-Server rather than the writer.
In the third scenario, processes initiate a non-
blocking pi_attach() on an inbound channel first,
before a pi_create(), followed by the write and
then read operations. This notifies the π-Server of
a pending request for a channel, providing writers
with the destination addresses and encouraging
Table 5. UDP request-reply performance within a cluster
Number of Clients
2
4
8
16
32
Number of Requests Served
80 000
160 000
320 000
640 000
1 280 000
Mean Execution Times (s)
21.99
25.51
38.72
44.65
96.58
Standard Deviation
0.822
0.813
0.741
0.430
0.170
Request Rate (per second)
3 642.6
6 277.6
8 267.1
14 335
13 390
Standard Deviation
138.89
200.85
162.38
137.58
23.82
Table 6. UDP request-reply performance on a WAN
Number of Clients
2
4
8
16
32
Number of Requests Served
10 000
20 000
40 000
80 000
160 000
Mean Execution Times (s)
10.58
11.02
11.52
12.34
24.74
Standard Deviation
0.518
0.340
0.285
0.122
0.268
Request Rate (per second)
946.4
1 815.6
3 473.8
6 480.8
6 467.6
Standard Deviation
46.32
56.53
86.53
64.07
70.39
Search WWH ::
Custom Search