The graph is of two test runs of the 3D simple test on the same machine. It shows the time taken to render each frame while the 3D test was running over a 9 second period. Run 1 (blue) was without having any network connections active. Run 2 (pink) was with a LAN card connected to the local network (but without any network I/O taking place).
It can be seen that there is background activity that occurs each and every second (the pink spikes) that causes the time taken to render a frame to blow out.
It is easy to disable a network connection before doing a benchmark test but similar effects have been noticed with other background tasks. Some of which are part of the O/S and can not be disabled without making the machine unstable.
A second effect can also be seen. There is a period of instability during the 1st second of the test. This is normal result of starting DirectX & generating textures used in the test.
On any Windows machine there are a host of other background activities that can be monitored. There are some well known offenders like antivirus programs, but many less well known ones as well. One detection method would be to use FileMon (from sysinternals) to monitor file accesses on a idle system. Doing this makes it easy to see some of the 'idle' activity. Some examples of things to be found include,
cisvc.exe - used to monitor the memory usage
cidaemon.exe - indexing service
explorer.exe - handling quotas
winlogin.exe - Windows Management Instrumentation (WMI)
svchost - More WMI
WMI (also known as Web-Based Enterprise Management (WBEM), is always running. It keeps a record of huge number of system statistics . e.g. CPU load, Network card bytes sent / received, I/O requests, cache performance, etc..
This is only the disk activity. There are other non disk activities as well. Mouse movements, time synchronisation, Windows update trying to connect to the net. Services starting and stopping (for example on my machine the IMAPI CD-Burning service starts and stops itself for no good reason).
Given PerformanceTest only runs for a short time even a quick interruption from these another tasks will impact the result to some degree.
Some of these backgound activities can be removed but many can't. But nevertheless it is good to be aware that they are there and take it into account when you are looking at benchmarks results.
-----
David