All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
3. Neither the name of the University nor the names of its contributors
may be used to endorse or promote products derived from this software
without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
SUCH DAMAGE.
@(#)2.t 5.1 (Berkeley) 4/17/91
Observation techniques
There are many tools available for monitoring the performance of the system. Those that we found most useful are described below. System maintenance tools
Several standard maintenance programs are invaluable in observing the basic actions of the system. The vmstat(1) program is designed to be an aid to monitoring systemwide activity. Together with the ps\|(1) command (as in ``ps av''), it can be used to investigate systemwide virtual memory activity. By running vmstat when the system is active you can judge the system activity in several dimensions: job distribution, virtual memory load, paging and swapping activity, disk and cpu utilization. Ideally, to have a balanced system in activity, there should be few blocked (b) jobs, there should be little paging or swapping activity, there should be available bandwidth on the disk devices (most single arms peak out at 25-35 tps in practice), and the user cpu utilization (us) should be high (above 50%).
If the system is busy, then the count of active jobs may be large, and several of these jobs may often be blocked (b). If the virtual memory is active, then the paging demon will be running (sr will be non-zero). It is healthy for the paging demon to free pages when the virtual memory gets active; it is triggered by the amount of free memory dropping below a threshold and increases its pace as free memory goes to zero.
If you run vmstat when the system is busy (a ``vmstat 5'' gives all the numbers computed by the system), you can find imbalances by noting abnormal job distributions. If many processes are blocked (b), then the disk subsystem is overloaded or imbalanced. If you have several non-dma devices or open teletype lines that are ``ringing'', or user programs that are doing high-speed non-buffered input/output, then the system time may go high (60-80% or higher). It is often possible to pin down the cause of high system time by looking to see if there is excessive context switching (cs), interrupt activity (in) or system call activity (sy). Long term measurements on one of our large machines show an average of 60 context switches and interrupts per second and an average of 90 system calls per second.
If the system is heavily loaded, or if you have little memory for your load (1 megabyte is little in our environment), then the system may be forced to swap. This is likely to be accompanied by a noticeable reduction in the system responsiveness and long pauses when interactive jobs such as editors swap out.
A second important program is iostat\|(1). Iostat iteratively reports the number of characters read and written to terminals, and, for each disk, the number of transfers per second, kilobytes transferred per second, and the milliseconds per average seek. It also gives the percentage of time the system has spent in user mode, in user mode running low priority (niced) processes, in system mode, and idling.
To compute this information, for each disk, seeks and data transfer completions and the number of words transferred are counted; for terminals collectively, the number of input and output characters are counted. Also, every 100 ms, the state of each disk is examined and a tally is made if the disk is active. From these numbers and the transfer rates of the devices it is possible to determine average seek times for each device.
When filesystems are poorly placed on the available disks, figures reported by iostat can be used to pinpoint bottlenecks. Under heavy system load, disk traffic should be spread out among the drives with higher traffic expected to the devices where the root, swap, and /tmp filesystems are located. When multiple disk drives are attached to the same controller, the system will attempt to overlap seek operations with I/O transfers. When seeks are performed, iostat will show non-zero average seek times. Most modern disk drives should exhibit an average seek time of 25-35 ms.
Terminal traffic reported by iostat should be heavily output oriented unless terminal lines are being used for data transfer by programs such as uucp. Input and output rates are system specific. Screen editors such as vi and emacs tend to exhibit output/input ratios of anywhere from 5/1 to 8/1. On one of our largest systems, 88 terminal lines plus 32 pseudo terminals, we observed an average of 180 characters/second input and 450 characters/second output over 4 days of operation. Kernel profiling
It is simple to build a 4.2BSD kernel that will automatically collect profiling information as it operates simply by specifying the -p option to config\|(8) when configuring a kernel. The program counter sampling can be driven by the system clock, or by an alternate real time clock. The latter is highly recommended as use of the system clock results in statistical anomalies in accounting for the time spent in the kernel clock routine.
Once a profiling system has been booted statistic gathering is handled by kgmon\|(8). Kgmon allows profiling to be started and stopped and the internal state of the profiling buffers to be dumped. Kgmon can also be used to reset the state of the internal buffers to allow multiple experiments to be run without rebooting the machine.
The profiling data is processed with gprof\|(1) to obtain information regarding the system's operation. Profiled systems maintain histograms of the kernel program counter, the number of invocations of each routine, and a dynamic call graph of the executing system. The postprocessing propagates the time spent in each routine along the arcs of the call graph. Gprof then generates a listing for each routine in the kernel, sorted according to the time it uses including the time of its call graph descendents. Below each routine entry is shown its (direct) call graph children, and how their times are propagated to this routine. A similar display above the routine shows how this routine's time and the time of its descendents is propagated to its (direct) call graph parents.
A profiled system is about 5-10% larger in its text space because of the calls to count the subroutine invocations. When the system executes, the profiling data is stored in a buffer that is 1.2 times the size of the text space. All the information is summarized in memory, it is not necessary to have a trace file being continuously dumped to disk. The overhead for running a profiled system varies; under normal load we see anywhere from 5-25% of the system time spent in the profiling code. Thus the system is noticeably slower than an unprofiled system, yet is not so bad that it cannot be used in a production environment. This is important since it allows us to gather data in a real environment rather than trying to devise synthetic work loads. Kernel tracing
The kernel can be configured to trace certain operations by specifying ``options TRACE'' in the configuration file. This forces the inclusion of code that records the occurrence of events in trace records in a circular buffer in kernel memory. Events may be enabled/disabled selectively while the system is operating. Each trace record contains a time stamp (taken from the VAX hardware time of day clock register), an event identifier, and additional information that is interpreted according to the event type. Buffer cache operations, such as initiating a read, include the disk drive, block number, and transfer size in the trace record. Virtual memory operations, such as a pagein completing, include the virtual address and process id in the trace record. The circular buffer is normally configured to hold 256 16-byte trace records.\** .FS \** The standard trace facilities distributed with 4.2 differ slightly from those described here. The time stamp in the distributed system is calculated from the kernel's time of day variable instead of the VAX hardware register, and the buffer cache trace points do not record the transfer size. .FE
Several user programs were written to sample and interpret the tracing information. One program runs in the background and periodically reads the circular buffer of trace records. The trace information is compressed, in some instances interpreted to generate additional information, and a summary is written to a file. In addition, the sampling program can also record information from other kernel data structures, such as those interpreted by the vmstat program. Data written out to a file is further buffered to minimize I/O load.
Once a trace log has been created, programs that compress and interpret the data may be run to generate graphs showing the data and relationships between traced events and system load.
The trace package was used mainly to investigate the operation of the file system buffer cache. The sampling program maintained a history of read-ahead blocks and used the trace information to calculate, for example, percentage of read-ahead blocks used. Benchmark programs
Benchmark programs were used in two ways. First, a suite of programs was constructed to calculate the cost of certain basic system operations. Operations such as system call overhead and context switching time are critically important in evaluating the overall performance of a system. Because of the drastic changes in the system between 4.1BSD and 4.2BSD, it was important to verify the overhead of these low level operations had not changed appreciably.
The second use of benchmarks was in exercising suspected bottlenecks. When we suspected a specific problem with the system, a small benchmark program was written to repeatedly use the facility. While these benchmarks are not useful as a general tool they can give quick feedback on whether a hypothesized improvement is really having an effect. It is important to realize that the only real assurance that a change has a beneficial effect is through long term measurements of general timesharing. We have numerous examples where a benchmark program suggests vast improvements while the change in the long term system performance is negligible, and conversely examples in which the benchmark program run more slowly, but the long term system performance improves significantly.