The following graph shows the cumulative distribution of machine idle times. This graph illustrates how much CPU capacity is available to be harnessed in our pool by scavenging idle cycles with Condor. Many of the workstations in our department's instructional labs are idle for less than 50% of the day, while many of the desktop workstations are idle for more than 75% of the day.
The following graph shows the cumulative distribution of machine idle periods (short, medium, and long). Idle periods less than 1 hour are considered short, periods between 1 and 3 hours are considered medium, and periods longer than 3 hours are considered long. The three percentages do not add up to 100% for each X value because the three datasets are sorted independently.
The following graph shows the cumulative distribution of total machine idle time divided into short, medium, and long idle periods. Since the percentages in this graph are based on the total time instead of the number of periods, the proportion of longer periods grows. Again, the three datasets are sorted independently. The length of machine idle times can give a sense of the importance of performing checkpoints at preemption time. The percentage of idle time for periods shorter than the periodic checkpointing interval (currently 3 hours in our pool) gives a sense of how much CPU capacity would be lost if checkpoints were not performed at preemption time. In fact, the actual amount lost would be greater than what is shown, since even for long idle periods the time since the last periodic checkpoint would be lost.
The following graph shows the cumulative distribution of idle periods by length. In other words, Y% of machine idle periods were of X or fewer hours in length. This graph confirms that about 50% of all idle periods are short (under 1 hour). We might expect clusters of idle periods around certain intervals, due to fixed interval activities which take users away from their keyboards (classes, meetings). For example, if classes are 50 minutes in length, we might expect to see a cluster of idle periods around 50 minutes, when users leave their machines to attend class. However, no such pattern appears for our department.
The following graph shows the cumulative distribution of total idle time by length of the idle period. In other words, Y% of machine idle time was obtained from idle periods of X or fewer hours in length.
CondorView Machine statistics pages provide additional information about the availability of machines in the pool. In particular, the CondorView graphs show how machine availability varies according to time of day and day of week.