“About the Process Accounting System” discusses the accounting subsystem that keeps track of system usage.
“Accounting Files and Directories” talks about where accounting information is stored.
“About Daily System Accounting” defines daily system accounting and gives setup procedures.
“Daily System Accounting with runacct” describes how to operate the runacct program.
The four initial sections describe the standard UNIX System V accounting procedures. IRIX also implements an extended accounting facility, discussed in the following section:
“IRIX Extended Accounting” describes an accounting subsystem useful for large computer sites.
The IRIX process accounting system can provide the following information:
The number of programs a user runs
The size and duration of user programs
Data throughput (I/O)
Using this information, you can:
Determine how system resources are used and if a particular user is using more than a reasonable share.
Trace significant system events, such as security breaches, by examining the list of all processes invoked by a particular user at a particular time.
Set up billing systems to charge login accounts for using system resources.
The next sections describe the parts of process accounting, how to turn on and off process accounting, and how to look at the various log files.
The IRIX process accounting system has several parts:
The IRIX kernel writes a record of each process on the system that terminates into the file /var/adm/pacct. The file contains one record per terminated process, organized according to the format defined in /usr/include/sys/acct.h.
You must specifically turn on this function. See “Turning On Process Accounting”.
Once process accounting is turned on, the cron program executes several accounting commands, as specified in /var/spool/cron/crontabs/adm and /var/spool/crontabs/root. The commands in adm perform monthly accounting (monacct), check the size of the pacct file (ckpacct), and provide a daily accounting of processes and connect time (runacct). The root crontab file runs the dodisk program, which provides a report on current disk usage. These commands run automatically when process accounting is turned on.
The login and init programs record connect sessions by writing records into /etc/wtmp. This happens by default, as long as the wtmp file exists.
Records of date changes, reboots, and shutdowns are copied from /etc/utmp to /etc/wtmp by the acctwtmp command.
The acctwtmp utility is automatically called by runacct, /usr/lib/acct/startacct, and /usr/lib/shutacct, once process accounting is turned on.
The disk utilization programs acctdusg and diskusg break down disk usage by login and prepare reports. These programs are run by the dodisk script. For details, see the acct(1M) , acctsh(1M) , and diskusg(1M) man pages.
Note that for XFS filesystems, disk quotas (installed with the subsystem eoe.sw.quotas) can be used as an efficient accounting tool to keep track of disk usage. Refer to IRIX Admin: Disks and Filesystems for more information.
To turn on process accounting:
Log in to the system as root.
Make sure the eoe.sw.acct subsystem is installed. If not, install it.
Enter this command:
chkconfig acct on
Enter this command:
This starts the kernel writing information into the file /var/adm/pacct.
Process accounting is started every time you boot the system, and every time the system boots, you should see a message similar to this:
System accounting started
Note that process accounting files, especially /var/adm/pacct, can grow very large. If you turn on process accounting, especially on a server, you should watch the amount of free disk space carefully. See “Accounting File Size Control”.
The directory /usr/lib/acct contains the programs and shell scripts necessary to run the accounting system. Process accounting uses a login (/var/adm) to perform certain tasks. /var/adm contains active data collection files used by the process accounting. Here is a description of the primary subdirectories in /var/adm:
/var/adm/acct/nite contains files that are reused daily by runacct.
/var/adm/acct/sum contains the cumulative summary files updated by runacct.
/var/adm/acct/fiscal contains periodic summary files created by monacct.
Process and disk accounting files can grow very large. On a busy system, they can grow quite rapidly.
To help keep the size of the file /var/adm/pacct under control, the cron command runs /usr/lib/acct/ckpacct to check the size of the file and the available disk space on the file system.
If the size of the pacct file exceeds 1000 blocks (by default), it runs the turnacct command with argument switch. The switch argument causes turnacct to back up the pacct file (removing any existing backup copy) and start a new, empty pacct file. This means that at any time, no more than 2000 blocks of disk space are taken by pacct file information.
If the amount of free space in the file system falls below 500 blocks, ckpacct automatically turns off process accounting by running the turnacct command with the off argument. When at least 500 blocks of disk space are free, accounting is activated again the next time cron runs ckpacct.
The files listed here are located in the /var/adm directory:
Diagnostic output during the execution of disk accounting programs
Output from the acctdusg program
Output from the chargefee program, ASCII tacct records
Active process accounting file
Process accounting files switched by turnacct
Process accounting files for MMDD during execution of runacct
The following files are located in the /var/adm/acct/nite directory:
Used by runacct to record progress and print warning and error messages. activeMMDD is the same as active after runacct detects an error .
ASCII total command summary used by prdaily.
Connect accounting records in tawcct.h format.
Output of acctcon1 program, connect session records in ctmp.h format.
ASCII daily command summary used by prdaily.
Total accounting records for one day in tacct.h format.
Disk accounting records in tacct.h format, created by dodisk procedure.
Diagnostic output during execution of runacct (see cron entry).
Last day runacct executed in date +%m%d format.
Used to control serial use of runacct.
tty line usage report used by prdaily.
Diagnostic output from acctcon1.
Same as log after runacct detects an error.
Contains beginning and ending dates from wtmp and contains a listing of reboots.
Used to record current state during execution of runacct.
wtmp file corrected by wtmpfix.
Place for wtmpfix error messages.
Same as wtmperror after runacct detects an error.
Previous day's wtmp file.
The following files are located in the /var/adm/acct/sum directory:
Total command summary file for current fiscal period in internal summary format
Command summary file without latest update
Command summary file for yesterday in internal summary format
Created by lastlogin
Concatenated version of all pacct files for MMDD, removed by remove procedure after reboot
Saved output of prdaily programs
Cumulative total accounting file for current fiscal period
Same as tacct without latest update
Total accounting file for MMDD
Saved copy of wtmp file for MMDD, removed by remove procedure after reboot
When IRIX enters multiuser mode, /usr/lib/acct/startup is executed as follows:
The acctwtmp program adds a boot record to /etc/wtmp. This record is signified by using the system name as the login name in the wtmp record.
Process accounting is started by turnacct, which in turn executes acct on /var/adm/pacct.
The remove command is executed to clean up the saved pacct and wtmp files left in the sum directory by runacct.
The ckpacct procedure is run through cron every hour of the day to check the size of /var/adm/pacct. If the file grows past 1000 blocks (default), the turnacct switch is executed. The advantage of having several smaller pacct files becomes apparent when you try to restart runacct after a failure processing these records.
The chargefee program can be used to bill users for file restores, and so on. It adds records to /var/adm/fee that are picked up and processed by the next execution of runacct and merged into the total accounting records. runacct is executed through cron each night. It processes the active accounting files, /var/adm/pacct, /etc/wtmp, /var/adm/acct/nite/disktacct, and /var/adm/fee. It produces command summaries and usage summaries by login name.
When the system is shut down using shutdown, the shutacct shell procedure is executed. It writes a shutdown reason record into /etc/wtmp and turns process accounting off.
After the first reboot each morning, the administrator should execute /usr/lib/acct/prdaily to print the previous day's accounting report.
If you have installed the system accounting option, all the files and command lines for implementation have been set up properly. You may wish to verify that the entries in the system configuration files are correct. In order to automate the operation of the accounting system, you should check that the following steps have been done:
The file /etc/init.d/acct should contain the following lines (among others):
The first line starts process accounting during the system startup process; the second stops it before the system is brought down.
For most installations, the following entries should be in /var/spool/cron/crontabs/adm so that cron automatically runs the daily accounting. These lines should already exist:
0 4 * * 1-6 if /etc/chkconfig acct; then /usr/lib/acct/runacct 2> /var/adm/acct/nite/fd2log; fi 5 * * * 1-6 if /etc/chkconfig acct; then /usr/lib/acct/ckpacct; fi
Note that the above cron commands must constitute only one line in the crontabs file. The following command, which also constitutes only one line in the crontabs file, should be in /var/spool/cron/crontabs/root:
0 2 * * 4 if /etc/chkconfig acct; then /usr/lib/acct/dodisk > /var/adm/acct/nite/disklog; fi
To facilitate monthly merging of accounting data, the following entry in /var/spool/cron/crontabs/adm allows monacct to clean up all daily reports and daily total accounting files, and deposit one monthly total report and one monthly total accounting file in the fiscal directory:
0 5 1 * * if /etc/chkconfig acct; then /usr/lib/acct/monacct; fi
The above command is all on one line in the source file, and takes advantage of the default action of monacct that uses the current month's date as the suffix for the file names. Notice that the entry is executed when runacct has sufficient time to complete. This will, on the first day of each month, create monthly accounting files with the entire month's data.
You may wish to verify that an account exists for adm. Also, verify that the PATH shell variable is set in /var/adm/.profile to:
To start up system accounting, simply type the commands
chkconfig acct on
The next time the system is booted, accounting starts.
The runacct command is the main daily accounting shell procedure. It is normally initiated by cron during nonpeak hours. runacct processes connect, fee, disk, and process accounting files. It also prepares daily and cumulative summary files for use by prdaily or for billing purposes.
The following files produced by runacct are of particular interest:
Produced by acctcon, reads the wtmp file and produces usage statistics for each terminal line on the system. This report is especially useful for detecting bad lines. If the ratio between the number of logoffs to logins exceeds about 3:1, it is quite possible that the line is failing.
The total accounting file for the previous day in tacct.h format.
The accumulation of each day's nite/daytacct can be used for billing purposes. It is restarted each month or fiscal period by the monacct procedure.
Produced by the acctcms program. It contains the daily command summary. The ASCII version of this file is nite/daycms.
The accumulation of each day's command summaries. It is restarted by the execution of monacct. The ASCII version is nite/cms.
Produced by the last login shell procedure. It maintains a record of the last time each login name was used.
Each execution of runacct saves a copy of the daily report that can be printed by prdaily.
The runacct command takes care not to damage files in the event of errors. A series of protection mechanisms are used that attempt to recognize an error, provide intelligent diagnostics, and terminate processing in such a way that runacct can be restarted with minimal intervention. It records its progress by writing descriptive messages into the file active. (Files used by runacct are assumed to be in the nite directory unless otherwise noted.) All diagnostics output during the execution of runacct are written into fd2log. runacct complains if the files lock and lockl exist when invoked. The lastdate file contains the month and day runacct was last invoked and is used to prevent more than one execution per day. If runacct detects an error, a message is written to /dev/console, mail is sent to root and adm, locks are removed, diagnostic files are saved, and execution is terminated.
To allow runacct to be restartable, processing is broken down into separate reentrant states. A file is used to remember the last state completed. When each state completes, statefile is updated to reflect the next state. After processing for the state is complete, statefile is read and the next state is processed. When runacct reaches the CLEANUP state, it removes the locks and terminates. States are executed as follows:
The command turnacct switch is executed. The process accounting files, /var/adm/pacct?, are moved to /var/adm/Spacct?.MMDD. The /etc/wtmp file is moved to /var/adm/acct/nite/wtmp.MMDD with the current time added on the end.
The wtmpfix program checks the wtmp file in the nite directory for correctness. Some date changes cause acctcon1 to fail, so wtmpfix attempts to adjust the time stamps in the wtmp file if a date change record appears.
Connect session records are written to ctmp in the form of ctmp.h. The lineuse file is created, and the reboots file is created showing all of the boot records found in the wtmp file.
The ctmp file is converted to ctacct.MMDD, which are connect accounting records. (Accounting records are in tacct.h format.)
The acctprc1 and acctprc2 programs are used to convert the process accounting files, /var/adm/Spacct?.MMDD, into total accounting records in ptacct?.MMDD. The Spacct and ptacct files are correlated by number so that if runacct fails, the unnecessary reprocessing of Spacct files will not occur. One precaution should be noted: when restarting runacct in this state, remove the last ptacct file, because it will not be complete.
Merge the process accounting records with the connect accounting records to form daytacct.
Merge in any ASCII tacct records from the file fee into daytacct.
On the day after the dodisk procedure runs, merge disktacct with daytacct.
Merge daytacct with sum/tacct, the cumulative total accounting file. Each day, daytacct is saved in sum/tacctMMDD, so that sum/tacct can be re-created in case it is corrupted or lost.
Merge in today's command summary with the cumulative command summary file sum/cms. Produce ASCII and internal format command summary files.
Any installation-dependent (local) accounting programs can be included here.
Clean up temporary files, run prdaily and save its output in sum/rprtMMDD, remove the locks, then exit.
The runacct procedure can fail for a variety of reasons—usually due to a system crash, /usr running out of space, or a corrupted wtmp file. If the activeMMDD file exists, check it first for error messages. If the active file and lock files exist, check fd2log for any mysterious messages. The following are error messages produced by runacct and the recommended recovery actions:
ERROR: locks found, run aborted
The files /var/adm/acct/nite/lock and /var/adm/acct/nite/lock1 were found. These files must be removed before runacct can restart.
ERROR: acctg already run for date: check /var/adm/acct/nite/lastdate
The date in lastdate and today's date are the same. Remove lastdate.
ERROR: turnacct switch returned rc=?
Check the integrity of turnacct and accton. The accton program must be owned by root and have the setuid bit set.
ERROR: Spacct?.MMDD already exists
File setups probably already run. Check status of files, then run setups manually.
ERROR: /var/adm/acct/nite/wtmp.MMDD already exists, run setup manually
ERROR: wtmpfix detected a corrupted wtmp file. Use fwtmp to correct the corrupted file.
ERROR: connect acctg failed: check /var/adm/acct/nite/log
The acctcon1 program encountered a bad wtmp file. Use fwtmp to correct the bad file.
ERROR: Invalid state, check /var/adm/acct/nite/active
The file statefile is probably corrupted. Check statefile for irregularities and read active before restarting.
The runacct program, called without arguments, assumes that this is the first invocation of the day. The argument MMDD is necessary if runacct is being restarted and specifies the month and day for which runacct will rerun the accounting. The entry point for processing is based on the contents of statefile. To override statefile, include the desired state on the command line. For example, to start runacct, use the command:
nohup runacct 2 /var/adm/acct/nite/fd2log &
To restart runacct:
nohup runacct 0601 2 /var/adm/acct/nite/fd2log &
To restart runacct at a specific state:
nohup runacct 0601 WTMPFIX 2 /var/adm/acct/nite/fd2log &
Sometimes, errors occur in the accounting system, and a file is corrupted or lost. You can ignore some of these errors, or simply restore lost or corrupted files from a backup. However, certain files must be fixed in order to maintain the integrity of the accounting system.
The wtmp files are the most delicate part of the accounting system. When the date is changed and the IRIX system is in multiuser mode, a set of date change records is written into /etc/wtmp. The wtmpfix program is designed to adjust the time stamps in the wtmp records when a date change is encountered. However, some combinations of date changes and reboots will slip through wtmpfix and cause acctcon1 to fail.
The following steps show how to fix a wtmp file:
fwtmp < wtmp.MMDD > xwtmp
Delete any corrupted records or delete all records from beginning up to the date change.
fwtmp -ic <wtmp> wtmp.MMDD
If the wtmp file is beyond repair, remove the file and create an empty wtmp file:
This prevents any charging of connect time. acctprc1 cannot determine which login owned a particular process, but it is charged to the login that is first in the password file for that user ID.
If the installation is using the accounting system to charge users for system resources, the integrity of sum/tacct is quite important. Occasionally, mysterious tacct records appear with negative numbers, duplicate user IDs, or a user ID of 65,535. First check sum/tacctprev with prtacct. If it looks all right, the latest sum/tacct.MMDD should be patched up, then sum/tacct re-created. A simple patchup procedure would be:
Enter the command:
Enter the command:
acctmerg -v < tacct.MMDD > xtacct
Enter the command:
Remove the bad records.
Write duplicate UID records to another file.
Enter the command:
acctmerg -i < xtacc t > tacct.MMDD
Enter the command:
acctmerg tacctprev <tacct.MMDD> tacct
Remember that the monacct procedure removes all the tacct.MMDD files; therefore, you can recreate sum/tacct by merging these files.
The file /usr/lib/acct/holidays contains the prime/nonprime table for the accounting system. The table should be edited to reflect your location's holiday schedule for the year. The format is composed of three types of entries:
Comment Lines, which may appear anywhere in the file as long as the first character in the line is an asterisk.
Year Designation Line, which should be the first data line (noncomment line) in the file and must appear only once. The line consists of three fields of four digits each (leading white space is ignored). For example, to specify the year as 1992, prime time at 9:00 a.m., and nonprime time at 4:30 p.m., the following entry is appropriate:
1992 0900 1630
A special condition allowed for in the time field is that the time 2400 is automatically converted to 0000.
Company Holidays Lines, which follow the year designation line and have the following general format:
day-of-year Month Day Description of Holiday
The day-of-year field is a number in the range of 1 through 366, indicating the day for the corresponding holiday (leading white space is ignored). The other three fields are actually commentary and are not currently used by other programs.
The runacct command generates five basic reports upon each invocation. They cover the areas of connect accounting, usage by person on a daily basis, command usage reported by daily and monthly totals, and a report of the last time users were logged in. The following paragraphs describe the reports and the meanings of their tabulated data.
In the first part of the report, the from/to banner should alert the administrator to the period reported on. This period runs from the time the last accounting report was generated until the time the current accounting report was generated. It is followed by a log of system reboots, shutdowns, power fail recoveries, and any other record dumped into /etc/wtmp by the acctwtmp program. See the acct(1M) man page for more information.
The second part of the report is a breakdown of line utilization. The TOTAL DURATION field tells how long the system was in multiuser state (able to be accessed through the terminal lines). The columns are:
The terminal line or access port.
The total number of minutes the line was in use during the accounting period.
The total number of minutes the line was in use divided into the total duration of the accounting period.
The number of times this port was accessed for a login session.
This column has little significance. It previously gave the number of times that the port was used to log a user on; but since login can no longer be executed explicitly to log in a new user, this column should be identical with SESS.
The number of times a user logged off and also any interrupts that occur on that line. Generally, interrupts occur on a port when getty is first invoked after the system is brought to a multiuser state. This column comes into play when the # OFF column exceeds the # ON column by a large factor. This usually indicates that the multiplexer, modem, or cable is going bad, or that there is a bad connection somewhere. The most common cause of this is an unconnected cable dangling from the multiplexer.
During real time, /etc/wtmp should be monitored, since this is the file from which connect accounting is geared. If it grows rapidly, execute acctcon1 to see which line is the noisiest. If the interrupting is occurring at a furious rate, general system performance will be affected.
The daily usage report gives a by-user breakdown of system resource utilization. Its data consists of:
The user ID.
The login name of the user; more than one login name can exist for a single user ID, and this entry identifies which login name used the resource.
The amount of time the user's process used the central processing unit. This category is broken down into PRIME and NPRIME (nonprime) utilization. The accounting system's idea of this breakdown is located in the /usr/lib/acct/holidays file. As delivered, prime time is defined to be 0900 through 1700 hours.
A cumulative measure of the amount of memory a process uses while running. The amount shown reflects kilobyte segments of memory used per minute. This measurement is also broken down into PRIME and NPRIME amounts.
The amount of time that a user was logged into the system. If this time is high and # OF PROCS is low, this indicates that the user was logged in for a long period of time without actually using the system. This column is also subdivided into PRIME and NPRIME utilization.
When the disk accounting programs have been run, the output is merged into the total accounting record (tacct.h) and shows up in this column. This disk accounting is accomplished by the program acctdusg.
|# OF PROCS|
The number of processes invoked by the user. Large numbers in this column indicate that a user may have had a shell running out of control.
|# O SESS|
Number of times the user logged onto the system.
|# DISK SAMPLES|
Number of times disk accounting was run to obtain the average number of DISK BLOCKS listed earlier.
An often unused field in the total accounting record, the FEE field represents the total accumulation of widgets charged against the user by the chargefee shell procedure. See the acctsh(1M) man page . The chargefee procedure is used to levy charges against a user for special services performed such as file restores, and so on.
The Daily Command and Monthly Total Command reports are virtually the same except that the Daily Command Summary reports only on the current accounting period, while the Monthly Total Command Summary tells the story for the start of the fiscal period to the current date. In other words, the monthly report reflects the data accumulated since the last invocation of monacct.
The data included in these reports tell an administrator which commands are used most heavily. Based on those commands' characteristics of system resource utilization, the administrator can decide what to weigh more heavily when system tuning. These reports are sorted by TOTAL KCOREMIN, which is an arbitrary yardstick but often a good one for calculating drain on a system.
The name of the command. Unfortunately, all shell procedures are lumped together under the name sh since only object modules are reported by the process accounting system. The administrator should monitor the frequency of programs called a.out or core or any other name that does not seem quite right. Often people like to work on their favorite version of a personal program, but they do not want everyone to know about it. acctcom is also a good tool for determining who executed a suspiciously named command and also to see if superuser privileges were abused.
The total number of invocations of this particular command.
The total cumulative measurement of the amount of kilobyte segments of memory used by a process per minute of run time.
The total processing time this program has accumulated.
The total real-time (wall-clock) minutes this program has accumulated. This total is the actual “waited for” time as opposed to kicking off a process in the background.
The mean of the TOTAL KCOREMIN over the number of invocations reflected by NUMBER CMDS.
The mean derived between NUMBER CMDS and TOTAL CPU-MIN.
This gives a relative measure of the total available CPU time consumed by the process during its execution. It is a measurement of the ratio of system availability to system utilization. It is computed by the formula:
total CPU time / elapsed time
This column, which may contain a negative value, is a total count of the number of characters pushed around by the read and write system calls.
A total count of the physical block reads and writes that a process performed.
Large computing sites often have many unrelated users and must be able to charge them separately for resource usage. Although there are IRIX mechanisms to provide usage information, these mechanisms are inadequate for many sites. Standard IRIX accounting lacks some important metrics, uses lots of disk space, and provides little flexibility for usage billing. Third-party accounting software addresses some of these issues, but is still limited by data provided by IRIX. Array clusters and hypercubes compound these problems by allowing a single user's resource usage to be spread over multiple systems.
IRIX provides three features to assist large computing sites with accounting needs: extended accounting, array sessions, and project IDs.
The original IRIX mechanism for resource accounting was based on standard UNIX System V accounting. Whenever a process exits, the kernel writes a record containing resource usage information to a file. Because the kernel itself does this file I/O, process accounting can become a minor bottleneck on heavily loaded systems. Another problem is with the format of data written by System V accounting: usage information is stored using an awkward compressed format that amounts to a 16-bit floating point number. Values quickly lose a significant amount of accuracy, and have a maximum value that is not difficult to exceed on modern systems (around 234, or 16 GB). Finally, there is no room for expansion in the accounting records. Metrics provided are fairly limited, and many customers need additional data. However, with no room for expansion, additional fields would require increasing the record size, which would break virtually all the existing software that uses accounting data.
In IRIX release 6.1 and later, extended accounting is available, while System V accounting remains in place, essentially unchanged.
One significant change in extended accounting is the delivery mechanism: records are written using the system audit trail (SAT) facility, which uses a daemon to collect audit records from the kernel using special system calls. SAT writes records out to destinations chosen by the system administrator; see the satd(1M) man page . This gets the kernel out of the file I/O business, and gives system administrators flexibility in the handling of accounting data.
The sat_select command can be used to select accounting events for the audit subsystem to monitor; see the sat_select(1M) man page for details.
Housekeeping tasks such as rotating audit files and handling file-system-full conditions are handled by the satd program. Third party software can either read the audit files in their entirety (files may contain records for non-accounting events if a site has elected to audit them) or use the existing sat_reduce program to filter out only relevant accounting records; see the sat_reduce(1M) man page . Contents of individual records can be dumped in ASCII format using the sat_interpret program; see the sat_interpret(1M) man page .
Resource data contained in extended accounting records is stored as uncompressed 64-bit values, which should be sufficient for most metrics into the near future. Records contain spare fields to allow for future expansion, and a version code to allow software to handle future format changes gracefully. In addition to all of the metrics reported by System V accounting, these new metrics have been added:
Number of swaps
Number of bytes read or written
Number of read or write requests
Time spent waiting
For block I/O
For physical I/O
On the run queue
To begin using extended accounting on a system, follow these steps:
Enable session accounting in the kernel by using the systune command to set the do_sessacct or do_extpacct parameters to nonzero values; see the systune(1M) man page .
Install the eoe.sw.audit subsystem from IRIX distribution media.
Enable the audit facility with the following command:
chkconfig audit on
Use the satconfig command to enable the sat_proc_acct or sat_session_acct audit events; see the satconfig(1M) man page . If you are using the audit facility for accounting purposes only, you may turn off all other events to conserve disk space.
For more information, see the extacct(1M) man page. Appendix A of IRIX Admin: System Configuration and Operation , lists kernel parameters for extended accounting.
To reduce disk space consumption and processing time for accounting records, IRIX can accumulate and report accounting information by array session. Process accounting is separately controlled— sites can use either accounting style, or both. Session accounting records contains data similar to process accounting records, except that counters and values reflect the accumulated total of all processes that were members of the session.
An array session is a set of processes all related to each other by a single unique identifier, the array session handle (ASH). A child process ordinarily inherits the ASH of its parent when created, thus becoming a member of its parent's array session. However, a system call is provided to allow a process to leave its parent's array session and start a new one. Programs like login and rshd use these system calls so that logging into the system effectively starts a new session. Programs like cron, su, and several batch queuing systems use these system calls so that work done on behalf of another user can have its own session. When the last process with a given ASH exits, the array sessions ends, and a session accounting record is written.
The ASH is a 64-bit value. A unique, increasing value (similar to a process ID) is assigned by default to each new array session as its handle. However, a system call is provided to change an array session's handle if desired. This can be used to synchronize the handles of array sessions on several systems in an array, thus allowing a distributed job to be considered a single entity for accounting purposes.
For more information, see the array_sessions(5) man page.
The range of handles that ths system assigns may be configurable, so it is possible to ensure that automatically assigned handles never conflict with process-specified ones. The system ensures that a particular ASH is never in use by more than one array session on that local system at one time.
In addition to accumulated totals of various process accounting data, session accounting records contain a 64-byte field intended for service provider information. In particular, batch queuing systems can use this field to record data about the queue name, initiator, and so forth. By default, the service provider information for a new array session is inherited from the array session of its parent process.
The standard init program always has its service provider information set to all zeros, and standard login utilities (login, su, rshd) never change service provider information. Batch queuing systems, on the other hand, are always expected to set service provider information to some nonzero value. Thus, it is possible to distinguish batch jobs from interactive sessions by checking if the service provider information is all zeros or not.
Many sites must be able to charge individual departments separately for their usage of a given system. Typically, this was done by billing total usage for each system user ID to the appropriate department. However, some sites have users that work for more than one department, so billing all usage to a single department is not appropriate.
To solve this accounting conundrum, the project ID feature was introduced into IRIX. A project ID is similar to a group ID, except that:
The current project ID is associated with an array session, not an individual process.
The project ID does not affect access permissions; its only purpose is for accounting.
A default project ID is associated with every user ID. Whenever it is necessary for a user to do work that should be billed to a different project, the newproj command may be used to change project ID; see the newproj(1) man page for details. This command starts a new shell and array session; background processes under the old shell continue being accounted for under the original project ID. Furthermore, the user ID and group ID remain unchanged, so access permissions are unaffected. To prevent users from specifying a project ID for which they are not authorized, the newproj command consults a file listing valid project IDs for each user. The system calls that set project ID require superuser privileges.
The file that contains user IDs and their authorized projects, /etc/project, is similar in style to /etc/passwd or /etc/group; see the project(4) man page for details. This file also specifies the default project for each user, in order to avoid modifying /etc/passwd. Because the project ID is a simple number, an additional file, /etc/projid, associates mnemonic ASCII names with numeric project IDs; see the projid(4) man page for details. The system administrator can configure a standard default project ID using the dfltprid variable of systune.
By default, an array session inherits the project ID of the session that spawned it. The standard login utilities (login, su, rshd) that start new array sessions have been updated to change project ID to the default project ID of the new user.
Library routines for reading project ID files is also provided, comparable to library routines for reading password file data. See the projid(3C) man page for more information.