The AVIDD-B and AVIDD-I Clusters
These two plots show the amount of CPUtime (#cpu*hours)
utilized per month.
These two plots show how much resource (#cpu*hours) were allocated to users.
This plot shows the difference between the amount of
resouces allocated to users and how much CPUtime
they actually used.
This plot shows the delivered TFlOPS for the AVIDD-IA32 clusters.
Notes:
CPUtime is derived from the systems' BSD Process Accounting records. CPUtime represents the convolution of system availability and the efficiency of user's code to consume CPU cycles. Simply put, if the systems are not available at all during a month, then utilization will be zero. Likewise, if users' code spend the month in a wait() state, then system utilization will also equal zero. To maximize this statistic, sys-admins need to maximize the availability of system resources and users need to maximize the amount of time their programs spend comsuming CPU cycles. Some things which adversly effect these numbers are system crashes, systems not being available for usage, users not utilizing all available resources, poor memory management, I/O wait states, and poor interprocess communications methods.
Allocated Time is derived from the PBS/Maui Resource Manager/Scheduler logs; most people will recognise this quanity as Walltime. PBS/Maui allocate resources to a single user; how they use those resources does not come into play with these statistics/plots. To maximize these statistics, it is encombant upon SysAdmins make as much resource (CPUs) available to users as possible and for users to use all the available resources. Some things which adversly effect these numbers are nodes offline, monthly maintenance activities, system crashes, etc.
The Comparison Plot shows the difference between what has been allocated to users and how much they actually use. It's purpose is to document inefficiences in user's codes; i.e. poor memory management, I/O wait states, and poor interprocess communications methods, etc.
The TFlOPS Plot scales the green bars above into units of total floating point operations delivered to users.




