Experiments

We choose to visualize the working sets of several applications that are representative of long-running computations on large data sets:

• LU, SP, MG, BT are from the NAS suite of benchmarks which are long-running applications on large data sets.
• N-BODY implements the Barnes-Hut simulation for the N-BODY problem, an application that seems to have no regular access pattern, since it traverses hierarchical data structures.
• GZIP is a file compressing utility, and GCC is a compiler, which are both representative of a program development environment.
• XLATR and ESIM are part of the CAD Benchmarks distributed by the Collaborative Benchmarking Laboratory (http://cbl.ncsu.edu/).
• The last three applications are from the database domain: OO7-T1 (part of the OO7 benchmark [1]) is a traversal of a database, OO7-Q7 is a query on the same database, and SQL is an SQL server running the TPC-B benchmark.

The characteristics of the applications are summarized in the following table:

We executed the above benchmark applications on top of the ATOM tracing environment [5] (apart from the SQL application for which we had the traces already available [4]), running on top of Digital UNIX V3.2D-1. To keep the tracing times reasonable, long running applications were run as long as necessary to collect one billion data references, which correspond to several billion instructions.

The working sets of the applications are shown in figures 1 to 12. Figures 1 to 9 have two views of the working set: The first one presents the memory access pattern of the application. This access pattern is calculated as follows: Time is divided in intervals (usually 2 million data references long). For each page accessed at least once, we plot a dot on the respective interval that the page was accessed. Thus, each dot in the figure corresponds to at least one access to a (4-KByte) page done sometime during a 2-million-references interval . For example, in figure 1 we see that most of the pages the LU application accesses range between 262000 and 268000. In most figures we also plot the size of the working set for each time interval.

  
Figure 1: Access Pattern and working set size of LU

  
Figure 2: Access Pattern and working set size of N-BODY simulation

  
Figure 3: Access Pattern and working set size of BT

  
Figure 4: Access Pattern and working set size of MG

  
Figure 5: Access Pattern and working set size of SP

  
Figure 6: Access Pattern and working set size of GCC

  
Figure 7: Access Pattern and working set size of GZIP

  
Figure 8: Access Pattern and working set size of XLATR

  
Figure 9: Access Pattern and working set size of ESIM

  
Figure 10: Access Pattern and working set size of OO7 T1

  
Figure 11: Access Pattern and working set size of OO7 Q7

  
Figure 12: Access Pattern of SQL