May 07, 2001, 12:56 PM —
Q: How can I tell which processes are causing problems and which ones are stuck in a bottleneck?
A: A significant amount of data is available that is not shown by the ps command. In
addition, there are more clever ways to process and display data than top or proctool use. A new extension to the SE toolkit implements some of my ideas in this area. Along the way it becomes clear that the CPU usage measurements everyone relies on are somewhat inaccurate.
Process data sources
I described process data sources in my August 1996 Performance Q&A
column, but this time I'll go a step further with the data.
These data structures are described in full in the proc(4)
manual page. They are also available in the SE toolkit, so if you want
to obtain the data and play around with it, you should look at the code
for ps-ax.se and msacct.se.
The interface to /proc involves sending ioctl
commands or opening special pseudo-files and reading them (a new
feature of Solaris 2.6). The data that ps uses is called
PIOCPSINFO. Here's what you get back from
ioctl (you get slightly different data if you read it from
the pseudo-file):
proc(4) File Formats proc(4)
PIOCPSINFO
This returns miscellaneous process information such as that
reported by ps(1). p is a pointer to a prpsinfo structure
containing at least the following fields:
typedef struct prpsinfo {
char pr_state; /* numeric process state (see pr_sname) */
char pr_sname; /* printable character representing pr_state */
char pr_zomb; /* !=0: process terminated but not waited for */
char pr_nice; /* nice for cpu usage */
u_long pr_flag; /* process flags */
int pr_wstat; /* if zombie, the wait() status */
uid_t pr_uid; /* real user id */
uid_t pr_euid; /* effective user id */
gid_t pr_gid; /* real group id */
gid_t pr_egid; /* effective group id */
pid_t pr_pid; /* process id */
pid_t pr_ppid; /* process id of parent */
pid_t pr_pgrp; /* pid of process group leader */
pid_t pr_sid; /* session id */
caddr_t pr_addr; /* physical address of process */
long pr_size; /* size of process image in pages */
long pr_rssize; /* resident set size in pages */
u_long pr_bysize; /* size of process image in bytes */
u_long pr_byrssize; /* resident set size in bytes */
caddr_t pr_wchan; /* wait addr for sleeping process */
short pr_syscall; /* system call number (if in syscall) */
id_t pr_aslwpid; /* lwp id of the aslwp; zero if no aslwp */
timestruc_t pr_start; /* process start time, sec+nsec since epoch */
timestruc_t pr_time; /* usr+sys cpu time for this process */
timestruc_t pr_ctime; /* usr+sys cpu time for reaped children */
long pr_pri; /* priority, high value is high priority */
char pr_oldpri; /* pre-SVR4, low value is high priority */
char pr_cpu; /* pre-SVR4, cpu usage for scheduling */
u_short pr_pctcpu; /* % of recent cpu time, one or all lwps */
u_short pr_pctmem; /* % of system memory used by the process */
dev_t pr_ttydev; /* controlling tty device (PRNODEV if none) */
char pr_clname[PRCLSZ]; /* scheduling class name */
char pr_fname[PRFNSZ]; /* last component of exec()ed pathname */
char pr_psargs[PRARGSZ];/* initial characters of arg list */
int pr_argc; /* initial argument count */
char **pr_argv; /* initial argument vector */
char **pr_envp; /* initial environment vector */
} prpsinfo_t;
You can get the data for each lightweight process of a multithreaded process separately.
