Memory Profiling

At the entry of each routine, TAU can inspect the heap memory used and trigger this value using user-defined events that tie it to the routine name. This generates performance data that includes statistics such as the number of samples, maximum, minimum, mean, and standard deviation of the value. The user must configure TAU with the -PROFILEMEMORY configuration option to record heap memory utilization in this manner. TAU also supports tracking memory headroom, or the amount of memory available on the heap before the program runs out of memory. Again, an event correlates this value with the routine name, and generates these statistics as shown in Figure 1. To enable collection of this information, the user must choose the -PROFILEHEADROOM option during configuration. On the Cray XT3, TAU uses the heap_info() system call to extract the heap memory utilization and memory headroom available. On other systems, it computes the headroom available using a series of malloc() calls that allocate chunks of memory. The amount of memory requested at each call is twice the amount requested in the previous call. When the system runs out of memory, it requests for the smallest chunk and if it successful in allocating it, it requests twice as much until it cannot allocate even the smallest amount. At that point, all memory is exhausted and TAU computes the memory headroom in this manner. Thereafter, it frees up all the memory blocks allocated in this manner and associates this data with the currently executing routine. Clearly, the presence of the heap_info() system call helps us develop memory introspection tools and it would help tool developers if such library routines were available on other systems as well. TAU uses system specific interfaces where available and portable mechanisms elsewhere.

TAU also supports tracking heap and headroom using timer interrupts for a single global event. Headroom can be associated with the currently executing callpath as well. However, these calls need special annotations in the source code, unlike the two configuration options described above.

Figure 3: Modifications to the application makefile to use TAU are minimal