Perform A Throughput Test; Perform A Latency Test - HP NC570C User Manual

72
# killall netserver
# netserver.sdp
# netperf.sdp -c -C -f g -H 192.168.0.2
Recv Send
Socket Socket
Size Size
bytes bytes
65535 65535
# netperf.sdp -c -C -f g -H 192.168.0.2 -t TCP_RR
TCP REQUEST/RESPONSE TEST to 192.168.0.2
Local /Remote
Socket Size
Send Recv
bytes bytes
65535 65535
10. Make sure processes include the SDP library when they load.
The /etc/ld.so.preload file tells the system's dynamic linker to load the SDP library when processes
are started.
a. Create the /etc/ld.so.preload file if the file does not exist.
b. Add the following line to /etc/ld.so.preload on both systems:
/lib/libsdp_sys.so
11. Stop the existing netserver daemon, which expects TCP connections over a normal network socket.
a. Stop the netserver daemon on Server 2, using the killall command.
Example

Perform a Throughput Test

12. Tell the SDP library that the next process should use SDP, and start the netserver process on
Server 2:
Example
13. Run the netperf SDP Throughput test on Server 1.
Example
Send
Message Elapsed
Size Time
bytes secs.
65535 10.00
14. Read the test results.
The throughput has increased about 50% from using IPoIB, and the CPU utilization has been
significantly reduced.

Perform a Latency Test

In addition to the Throughput test, you can also test the effect of using SDP on the request/response test:
Example
Request Resp. Elapsed Trans.
Size Size Time
bytes bytes secs.
1 1 10.00
In the example above, there is approximately a 50% increase in transactions per second from the
IPoIB case. In addition, there is a reduction in CPU utilization on both the transmit and receive end.
Utilization
Send
Throughput local
10^9bits/s % T
1.94 36.70
CPU
Rate local
per sec % T
17145.82 15.00
Service Demand
Recv Send
remote local
% T us/KB
57.90 1.551
CPU S.dem S.dem
remote local remote
% T us/Tr us/Tr
17.10 8.747
Recv
remote
us/KB
2.446
9.976