What Data the Collector Collects

The Collector collects three different kinds of data: profiling data, tracing data and global data.

• Profiling data is collected by recording profile events at regular intervals. The interval is either a time interval obtained by using the system clock or a number of hardware events of a specific type. When the interval expires, a signal is delivered to the system and the data is recorded at the next opportunity.

• Tracing data is collected by interposing a wrapper function on various system functions so that calls to the system functions can be intercepted and data recorded about the calls.

• Global data is collected by calling various system routines to obtain information. The global data packet is called a sample.

Both profiling data and tracing data contain information about specific events, and both types of data are converted into performance metrics. Global data is not converted into metrics, but is used to provide markers that can be used to divide the program execution into time segments. The global data gives an overview of the program execution during that time segment.

The data packets collected at each profiling event or tracing event include the following information:

• A header identifying the data

• A high-resolution timestamp

• A thread ID

• A lightweight process (LWP) ID

• A processor (CPU) ID, when available from the operating system

• A copy of the call stack. For Java programs, two call stacks are recorded: the machine call stack and the Java call stack.

• For OpenMP programs, an identifier for the current parallel region and the OpenMP state are also collected.

For more information on threads and lightweight processes, see Chapter 7, Understanding the Performance Analyzer and Its Data.

In addition to the common data, each event-specific data packet contains information specific to the data type. The five types of data that the Collector can record are:

• Clock profile data

• Hardware counter overflow profiling data

• Synchronization wait tracing data

• Heap tracing (memory allocation) data

• MPI tracing data

These five data types, the metrics that are derived from them, and how you might use them, are described in the following subsections. A sixth type of data, global sampling data, cannot be converted to metrics because it does not include call stack information.

Clock Data

When you are doing clock-based profiling, the data collected depends on the metrics provided by the operating system.

Clock-based Profiling Under the Solaris OS

In clock-based profiling under the Solaris OS, the state of each LWP is stored at regular time intervals. This time interval is called the profiling interval. The information is stored in an integer array: one element of the array is used for each of the ten microaccounting states maintained by the kernel. The data collected is converted by the Performance Analyzer into times spent in each state, with a resolution of the profiling interval. The default profiling interval is approximately 10 milliseconds (10 ms). The Collector provides a high-resolution profiling interval of approximately 1 ms and a low-resolution profiling interval of approximately 100 ms, and, where the OS permits, allows arbitrary intervals. Running the collect command with no arguments prints the range and resolution allowable on the system on which it is run.

The metrics that are computed from clock-based data are defined in the following table.

For multithreaded experiments, times other than wall clock time are summed across all LWPs. Wall time as defined is not meaningful for multiple-program multiple-data (MPMD) programs.

Timing metrics tell you where your program spent time in several categories and can be used to improve the performance of your program.

• High user CPU time tells you where the program did most of the work. It can be used to find the parts of the program where there may be the most gain from redesigning the algorithm.

• High system CPU time tells you that your program is spending a lot of time in calls to system routines.

• High wait CPU time tells you that there are more threads ready to run than there are CPUs available, or that other processes are using the CPUs.

• High user lock time tells you that threads are unable to obtain the lock that they request.

• High text page fault time means that the code generated by the linker is organized in memory so that calls or branches cause a new page to be loaded. Creating and using a mapfile (see “Generating and Using a Mapfile” in the Performance Analyzer online help) can fix this kind of problem.

• High data page fault time indicates that access to the data is causing new pages to be loaded. Reorganizing the data structure or the algorithm in your program can fix this problem.

Clock-based Profiling Under the Linux OS

Under the Linux OS, the only metric available is User CPU time. Although the total CPU utilization time reported is accurate, it may not be possible for the Analyzer to determine the proportion of the time that is actually System CPU time as accurately as for the Solaris OS. Although the Analyzer displays the information as if the data were for a lightweight process (LWP), in reality there are no LWP’s on a Linux OS; the displayed LWP ID is actually the thread ID.

Clock-based Profiling for MPI Programs

When clock-profiling data is collected on an MPI experiment run with a version of Sun HPC ClusterTools that has functionality for MPI State profiling, two additional metrics can be shown:

• MPI Work, which accumulates when the process is inside the MPI runtime doing work, such as processing requests or messages

• MPI Wait, which accumulates when the process is inside the MPI runtime, but waiting for an event, buffer, or message

On the Linux OS, MPI Work and MPI Wait are accumulated only when the process is active in either user or system mode. Unless you have specified that MPI should do a busy wait, MPI Wait on Linux is not useful.

Clock-based Profiling for OpenMP Programs

If clock-based profiling is performed on an OpenMP program, two additional metrics are provided: OpenMP Work and OpenMP Wait.

On the Solaris OS, OpenMP Work accumulates when work is being done either serially or in parallel. OpenMP Wait accumulates when the OpenMP runtime is waiting for synchronization, and accumulates whether the wait is using CPU time or sleeping, or when work is being done in parallel, but the thread is not scheduled on a CPU.

On the Linux OS, OpenMP Work and OpenMP Wait are accumulated only when the process is active in either user or system mode. Unless you have specified that OpenMP should do a busy wait, OpenMP Wait on Linux is not useful.

Hardware Counter Overflow Profiling Data

Hardware counters keep track of events like cache misses, cache stall cycles, floating-point operations, branch mispredictions, CPU cycles, and instructions executed. In hardware counter overflow profiling, the Collector records a profile packet when a designated hardware counter of the CPU on which an LWP is running overflows. The counter is reset and continues counting. The profile packet includes the overflow value and the counter type.

Various CPU families support from two to eighteen simultaneous hardware counter registers. The Collector can collect data on one or more registers. For each register the Collector allows you to select the type of counter to monitor for overflow, and to set an overflow value for the counter. Some hardware counters can use any register, others are only available on a particular register. Consequently, not all combinations of hardware counters can be chosen in a single experiment.

Hardware counter overflow profiling data is converted by the Performance Analyzer into count metrics. For counters that count in cycles, the metrics reported are converted to times; for counters that do not count in cycles, the metrics reported are event counts. On machines with multiple CPUs, the clock frequency used to convert the metrics is the harmonic mean of the clock frequencies of the individual CPUs. Because each type of processor has its own set of hardware counters, and because the number of hardware counters is large, the hardware counter metrics are not listed here. The next subsection tells you how to find out what hardware counters are available.

One use of hardware counters is to diagnose problems with the flow of information into and out of the CPU. High counts of cache misses, for example, indicate that restructuring your program to improve data or text locality or to increase cache reuse can improve program performance.

Some of the hardware counters correlate with other counters. For example, branch mispredictions and instruction cache misses are often related because a branch misprediction causes the wrong instructions to be loaded into the instruction cache, and these must be replaced by the correct instructions. The replacement can cause an instruction cache miss, or an instruction translation lookaside buffer (ITLB) miss, or even a page fault.

Hardware counter overflows are often delivered one or more instructions after the instruction which caused the event and the corresponding event counter to overflow: this is referred to as “skid” and it can make counter overflow profiles difficult to interpret. In the absence of hardware support for precise identification of the causal instruction, an apropos backtracking search for a candidate causal instruction may be attempted.

When such backtracking is supported and specified during collection, hardware counter profile packets additionally include the PC (program counter) and EA (effective address) of a candidate memory-referencing instruction appropriate for the hardware counter event. (Subsequent processing during analysis is required to validate the candidate event PC and EA.) This additional information about memory-referencing events facilitates various data-oriented analyses. Backtracking is supported only on SPARC based platforms running the Solaris OS.

Backtracking and recording of a candidate event PC and EA can also be specified for clock-profiling, although it might be difficult to interpret.

Hardware Counter Lists

Hardware counters are processor-specific, so the choice of counters available to you depends on the processor that you are using. The performance tools provide aliases for a number of counters that are likely to be in common use. You can obtain a list of available hardware counters on any particular system from the Collector by typing collect with no arguments in a terminal window on that system. If the processor and system support hardware counter profiling, the collect command prints two lists containing information about hardware counters. The first list contains hardware counters that are aliased to common names; the second list contains raw hardware counters. If neither the performance counter subsystem nor the collect command know the names for the counters on a specific system, the lists are empty. In most cases, however, the counters can be specified numerically.

Here is an example that shows the entries in the counter list. The counters that are aliased are displayed first in the list, followed by a list of the raw hardware counters. Each line of output in this example is formatted for print.

Aliased HW counters available for profiling:
cycles[/{0|1}],9999991 (’CPU Cycles’, alias for Cycle_cnt; CPU-cycles)
insts[/{0|1}],9999991 (’Instructions Executed’, alias for Instr_cnt; events)
dcrm[/1],100003 (’D$ Read Misses’, alias for DC_rd_miss; load events)
...
Raw HW counters available for profiling:
Cycle_cnt[/{0|1}],1000003 (CPU-cycles)
Instr_cnt[/{0|1}],1000003 (events)
DC_rd[/0],1000003 (load events)

Format of the Aliased Hardware Counter List

In the aliased hardware counter list, the first field (for example, cycles) gives the alias name that can be used in the -h counter... argument of the collect command. This alias name is also the identifier to use in the er_print command.

The second field lists the available registers for the counter; for example, [/{0|1}].

The third field, for example, 9999991, is the default overflow value for the counter. For aliased counters, the default value has been chosen to provide a reasonable sample rate. Because actual rates vary considerably, you might need to specify a non-default value.

The fourth field, in parentheses, contains type information. It provides a short description (for example, CPU Cycles), the raw hardware counter name (for example, Cycle_cnt), and the type of units being counted (for example, CPU-cycles).

If the first word of type information is:

• load, store, or load-store, the counter is memory-related. You can prepend a + sign to the counter name (for example, +dcrm) in the collect -h command, to request a search for the precise instruction and virtual address that caused the event. The + sign also enables dataspace profiling; see The DataObjects Tab, The DataLayout Tab, and The MemoryObjects Tabs for details.

• not-program-related, the counter captures events initiated by some other program, such as CPU-to-CPU cache snoops. Using the counter for profiling generates a warning and profiling does not record a call stack.

If the second or only word of the type information is:

• CPU-cycles, the counter can be used to provide a time-based metric. The metrics reported for such counters are converted by default to inclusive and exclusive times, but can optionally be shown as event counts.

• events, the metric is inclusive and exclusive event counts, and cannot be converted to a time.

In the aliased hardware counter list in the example, the type information contains one word, CPU-cycles for the first counter and events for the second counter. For the third counter, the type information contains two words, load events.

Format of the Raw Hardware Counter List

The information included in the raw hardware counter list is a subset of the information in the aliased hardware counter list. Each line in the raw hardware counter list includes the internal counter name as used by cpu-track(1), the register numbers on which that counter can be used, the default overflow value, and the counter units, which can be either CPU-cycles or Events.

If the counter measures events unrelated to the program running, the first word of type information is not-program-related. For such a counter, profiling does not record a call stack, but instead shows the time being spent in an artificial function, collector_not_program_related . Thread and LWP ID’s are recorded, but are meaningless.

The default overflow value for raw counters is 1000003. This value is not ideal for most raw counters, so you should specify overflow values when specifying raw counters.

Synchronization Wait Tracing Data

In multithreaded programs, the synchronization of tasks performed by different threads can cause delays in execution of your program, because one thread might have to wait for access to data that has been locked by another thread, for example. These events are called synchronization delay events and are collected by tracing calls to the Solaris or pthread thread functions. The process of collecting and recording these events is called synchronization wait tracing. The time spent waiting for the lock is called the wait time. Currently, synchronization wait tracing is only available for systems running the Solaris OS.

Events are only recorded if their wait time exceeds a threshold value, which is given in microseconds. A threshold value of 0 means that all synchronization delay events are traced, regardless of wait time. The default threshold is determined by running a calibration test, in which calls are made to the threads library without any synchronization delay. The threshold is the average time for these calls multiplied by an arbitrary factor (currently 6). This procedure prevents the recording of events for which the wait times are due only to the call itself and not to a real delay. As a result, the amount of data is greatly reduced, but the count of synchronization events can be significantly underestimated.

Synchronization tracing is not supported for Java programs.

Synchronization wait tracing data is converted into the following metrics.

From this information you can determine if functions or load objects are either frequently blocked, or experience unusually long wait times when they do make a call to a synchronization routine. High synchronization wait times indicate contention among threads. You can reduce the contention by redesigning your algorithms, particularly restructuring your locks so that they cover only the data for each thread that needs to be locked.

Heap Tracing (Memory Allocation) Data

Calls to memory allocation and deallocation functions that are not properly managed can be a source of inefficient data usage and can result in poor program performance. In heap tracing, the Collector traces memory allocation and deallocation requests by interposing on the C standard library memory allocation functions malloc, realloc, valloc, and memalign and the deallocation function free. Calls to mmap are treated as memory allocations, which allows heap tracing events for Java memory allocations to be recorded. The Fortran functions allocate and deallocate call the C standard library functions, so these routines are traced indirectly.

Heap profiling for Java programs is not supported.

Heap tracing data is converted into the following metrics.

Collecting heap tracing data can help you identify memory leaks in your program or locate places where there is inefficient allocation of memory.

Another definition of memory leaks that is commonly used, such as in the dbx debugging tool, says a memory leak is a dynamically-allocated block of memory that has no pointers pointing to it anywhere in the data space of the program. The definition of leaks used here includes this alternative definition, but also includes memory for which pointers do exist.

MPI Tracing Data

The Collector can collect data on calls to the Message Passing Interface (MPI) library.

MPI tracing is implemented using the open source VampirTrace 5.5.3 release. It recognizes the following VampirTrace environment variables:

For more information on these variables, see the Vampirtrace User Manual on the Technische Universität Dresden web site.

MPI events that occur after the buffer limits have been reached are not written into the trace file resulting in an incomplete trace.

To remove the limit and get a complete trace of an application, set the VT_MAX_FLUSHES environment variable to 0. This setting causes the MPI API trace collector to flush the buffer to disk whenever the buffer is full.

To change the size of the buffer, set the VT_BUFFER_SIZE environment variable. The optimal value for this variable depends on the application that is to be traced. Setting a small value increases the memory available to the application, but triggers frequent buffer flushes by the MPI API trace collector. These buffer flushes can significantly change the behavior of the application. On the other hand, setting a large value such as 2G minimizes buffer flushes by the MPI API trace collector, but decreases the memory available to the application. If not enough memory is available to hold the buffer and the application data, parts of the application might be swapped to disk leading to a significant change in the behavior of the application.

The functions for which data is collected are listed below.

MPI tracing data is converted into the following metrics.

MPI Time is the total LWP time spent in the MPI function. If MPI state times are also collected, MPI Work Time plus MPI Wait Time for all MPI functions other than MPI_Init and MPI_Finalize should approximately equal MPI Work Time. On Linux, MPI Wait and Work are based on user+system CPU time, while MPI Time is based on real tine, so the numbers will not match.

MPI byte and message counts are currently collected only for point‐to‐point messages; they are not recorded for collective communication functions. The MPI Bytes Received metric counts the actual number of bytes received in all messages. MPI Bytes Sent counts the actual number of bytes sent in all messages. MPI Sends counts the number of messages sent, and MPI Receives counts the number of messages received.

Collecting MPI tracing data can help you identify places where you have a performance problem in an MPI program that could be due to MPI calls. Examples of possible performance problems are load balancing, synchronization delays, and communications bottlenecks.

Global (Sampling) Data

Global data is recorded by the Collector in packets called sample packets. Each packet contains a header, a timestamp, execution statistics from the kernel such as page fault and I/O data, context switches, and a variety of page residency (working-set and paging) statistics. The data recorded in sample packets is global to the program and is not converted into performance metrics. The process of recording sample packets is called sampling.

Sample packets are recorded in the following circumstances:

• When the program stops for any reason in the Debugging window of the IDE or in dbx, such as at a breakpoint, if the option to do this is set

• At the end of a sampling interval, if you have selected periodic sampling. The sampling interval is specified as an integer in units of seconds. The default value is 1 second

• When you choose Advanced -> Advanced Profiling -> Configure in the IDE and select the Periodic Samples checkbox in the Project Properties dialog box

• When you use the dbx collector sample record command to manually record a sample

• At a call to collector_sample, if you have put calls to this routine in your code (see Program Control of Data Collection)

• When a specified signal is delivered, if you have used the -l option with the collect command (see the collect(1) man page)

• When collection is initiated and terminated

• When you pause collection with the dbx collector pause command (just before the pause) and when you resume collection with the dbx collector resume command (just after the resume)

• Before and after a descendant process is created

The performance tools use the data recorded in the sample packets to group the data into time periods, which are called samples. You can filter the event-specific data by selecting a set of samples, so that you see only information for these particular time periods. You can also view the global data for each sample.

The performance tools make no distinction between the different kinds of sample points. To make use of sample points for analysis you should choose only one kind of point to be recorded. In particular, if you want to record sample points that are related to the program structure or execution sequence, you should turn off periodic sampling, and use samples recorded when dbx stops the process, or when a signal is delivered to the process that is recording data using the collect command, or when a call is made to the Collector API functions.