VMware vSphere 4, Architecture And Performance

Page: 9 / 19

9
VMware white paper
3.3. Netperf Throughput
Netperf is a micro-benchmark that measures the throughput of sending and receiving network packets. In this experiment netperf
was configured so packets could be sent continuously without having to wait for acknowledgements. Since all the receive traffic
needs to be recorded and then transmitted to the secondary, netperf Rx represents a workload with significant FT logging traffic.
As shown inFigure 6, with FT enabled, the virtual machine received 890 Mbits/sec of traffic while generating and sending 950 Mbits/
sec of logging traffic to the secondary. Transmit traffic, on the other hand, produced relatively little FT logging traffic, mostly from
acknowledgement responses and transmit completion interrupt events.
Figure 6. Netperf Throughput
0
100
200
300
500
400
600
700
800
900
1000
FT Disabled
FT Enabled
Netperf throughput
Mbits/sec
Receives Transmits
FT traffic Rx: 950 Mbits/sec
FT traffic Tx: 54 Mbits/sec
3.4. Netperf Latency Bound Case
In this experiment, netperf was configured to use the same message and socket size so that outstanding messages could only be
sent one at a time. Under this setup, the TCP/IP stack of the sender has to wait for an acknowledgment response from the receiver
before sending the next message and, thereby, any increase in latency results in a corresponding drop in network throughput.
Note in reality almost all applications send multiple messages without waiting for acknowledgement, so application throughput is
not impacted by any increase in network latency. However since this experiment was purposefully designed to test the worst-case
scenario, the throughput was made dependent on the network latency. There are not any known real world applications that would
exhibit this behavior. As discussed in
section 1.4 , when FT is enabled, the primary virtual machine delays the network transmit until
the secondary acknowledges that it has received all the events preceding the transmission of that packet. As expected, FT enabled
virtual machines had higher latencies which caused a corresponding drop in throughput.