ts_dptbl(4) ts_dptbl(4)
NAME
tsdptbl - time-sharing dispatcher parameter table
DESCRIPTION
The process scheduler (or dispatcher) is the portion of the kernel
that controls allocation of the CPU to processes. The scheduler sup-
ports the notion of scheduling classes where each class defines a
scheduling policy, used to schedule processes within that class. Asso-
ciated with each scheduling class is a set of priority queues on which
ready to run processes are linked. These priority queues are mapped by
the system configuration into a set of global scheduling priorities
which are available to processes within the class. (The dispatcher
always selects for execution the process with the highest global
scheduling priority that is eligible to run on a given CPU. See PRO-
CESSOR AFFINITY AND BINDING CONSIDERATIONS below.) The priority queues
associated with a given class are viewed by that class as a contiguous
set of priority levels numbered from 0 (lowest priority) to n (highest
priority - a configuration-dependent value). The set of global
scheduling priorities that the queues for a given class are mapped
into might not start at zero and might not be contiguous (depending on
the configuration).
Processes in the time-sharing class which are running in user mode (or
in kernel mode before going to sleep) are scheduled according to the
parameters in a time-sharing dispatcher parameter table (tsdptbl).
(Time-sharing processes running in kernel mode after sleeping are run
within a special range of priorities reserved for such processes and
are not affected by the parameters in the tsdptbl until they return
to user mode.) The tsdptbl consists of an array of parameter struc-
tures (struct tsdpent), one for each of the n priority levels used by
time-sharing processes in user mode. The properties of a given prior-
ity level i are specified by the ith parameter structure in this array
(tsdptbl[i]).
A parameter structure consists of the following members. These are
also described in the /usr/include/sys/ts.h header file.
tsglobpri The global scheduling priority associated with this
priority level. The mapping between time-sharing prior-
ity levels and global scheduling priorities is deter-
mined at boot time by the system configuration.
tsglobpri is the only member of the tsdptbl which
cannot be changed with dispadmin(1M).
tsquantum The length of the time quantum allocated to processes
at this level in ticks (HZ).
tstqexp Priority level of the new queue on which to place a
process running at the current level if it exceeds its
time quantum. Normally this field links to a lower
priority time-sharing level that has a larger quantum.
Page 1 Reliant UNIX 5.44 Printed 11/98
ts_dptbl(4) ts_dptbl(4)
tsslpret Priority level of the new queue on which to place a
process, that was previously in user mode at this
level, when it returns to user mode after sleeping.
Normally this field links to a higher priority level
that has a smaller quantum.
tsmaxwait A per process counter, tsdispwait is initialized to
zero each time a time-sharing process is placed back on
the dispatcher queue after its time quantum has expired
or when it is awakened (tsdispwait is not reset to
zero when a process is preempted by a higher priority
process). This counter is incremented once per second
for each process on the dispatcher queue. If a
process's tsdispwait value exceeds the tsmaxwait
value for its level, the process's priority is changed
to that indicated by tslwait. The purpose of this
field is to prevent starvation.
tslwait Move a process to this new priority level if
tsdispwait is greater than tsmaxwait.
An administrator can affect the behavior of the time-sharing portion
of the scheduler by reconfiguring the tsdptbl. There are two methods
available for doing this.
KERNEL LINK TIME CONFIGURATION
The tsdptbl can be reconfigured at kernel link time time by specify-
ing the desired values in the /etc/conf/pack.d/ts/space.c file and
reconfiguring the system using the auto-configuration procedure [see
idbuild(1M)]. This is the only method that can be used to change the
number of time share priority levels or the set of global scheduling
priorities used by the time share class.
DISPADMIN CONFIGURATION FILE
With the exception of tsglobpri all of the members of the tsdptbl
can be examined and modified on a running system using the
dispadmin(1M) command. Invoking dispadmin for the time-sharing class
allows the administrator to retrieve the current tsdptbl configura-
tion from the kernel's in-core table, or overwrite the in-core table
with values from a configuration file. The configuration file used for
input to dispadmin must conform to the specific format described
below.
Blank lines are ignored and any part of a line to the right of a #
symbol is treated as a comment. The first non-blank, non-comment line
must indicate the resolution to be used for interpreting the
tsquantum time quantum values. The resolution is specified as
RES=res
Page 2 Reliant UNIX 5.44 Printed 11/98
ts_dptbl(4) ts_dptbl(4)
where res is a positive integer between 1 and 1,000,000,000 inclusive
and the resolution used is the reciprocal of res in seconds (for exam-
ple, RES=1000 specifies millisecond resolution). Although very fine
(nanosecond) resolution may be specified, the time quantum lengths are
rounded up to the next integral multiple of the system clock's resolu-
tion.
The remaining lines in the file are used to specify the parameter
values for each of the time-sharing priority levels. The first line
specifies the parameters for time-sharing level 0, the second line
specifies the parameters for time-sharing level 1, etc. There must be
exactly one line for each configured time-sharing priority level.
EXAMPLE
The following excerpt from a dispadmin configuration file illustrates
the format. Note that for each line specifying a set of parameters
there is a comment indicating the corresponding priority level. These
level numbers indicate priority within the time-sharing class, and the
mapping between these time-sharing priorities and the corresponding
global scheduling priorities is determined by the configuration speci-
fied in the ts/space.c file. The level numbers are strictly for the
convenience of the administrator reading the file and, as with any
comment, they are ignored by dispadmin. dispadmin assumes that the
lines in the file are ordered by consecutive, increasing priority
level (from 0 to the maximum configured time-sharing priority). The
level numbers in the comments should normally agree with this order-
ing; if for some reason they don't, however, dispadmin is unaffected.
# Time Sharing Dispatcher Configuration
RES=1000
# tsquantum tstqexp tsslpret tsmaxwait tslwait PRIORITY LEVEL
1000 0 10 5 10 # 0
1000 0 11 5 11 # 1
1000 1 12 5 12 # 2
1000 1 13 5 13 # 3
1000 2 14 5 14 # 4
1000 2 15 5 15 # 5
1000 3 16 5 16 # 6
1000 3 17 5 17 # 7
1000 4 18 5 18 # 8
1000 4 19 5 19 # 9
800 5 20 5 20 # 10
800 5 21 5 21 # 11
. . . . . . .
100 48 59 5 59 # 58
100 49 59 5 59 # 59
Page 3 Reliant UNIX 5.44 Printed 11/98
ts_dptbl(4) ts_dptbl(4)
PROCESSOR AFFINITY AND BINDING CONSIDERATIONS
In order to increase the cache hit rates of a multiprocessor, a pro-
cess should generally execute on a particular processor. This affinity
for a particular processor will reduce the required memory bandwidth
for the process (and for the system as a whole). If a process has not
run "recently" then it probably only has a small cache state in any
given processor - and thus should loose its affinity for any particu-
lar processor.
This affinity mechanism can be enabled or disabled on a per-process
basis by the super user [see mpcntl(1M), mpcntl(3X)]. The affinity
mechanism defaults to on for the TS class.
A process can also be bound to a set of processors. If a processor has
been so bound, then it will only run on one of the processors of that
set. If there is a idle processor that is not in the processes binding
set - then the process will still wait until there is a processor in
its set that is not running a higher priority process.
Process bindings can be established by the process [see mpcntl(3X)],
or can be temporarily established by the kernel (if the kernel is exe-
cuting driver code that must run on a particular CPU).
NOTES
dispadmin does some limited sanity checking on the values supplied in
the configuration file. The sanity checking is intended to ensure that
the new tsdptbl values do not cause the system to panic. The sanity
checking does not attempt to analyze the effect that the new values
will have on the performance of the system. Unusual tsdptbl confi-
gurations may have a dramatic negative impact on the performance of
the system.
No sanity checking is done on the tsdptbl values specified in the
ts/space.c file. Specifying an inconsistent or nonsensical tsdptbl
configuration through the ts/space.c file could cause serious perfor-
mance problems and/or cause the system to panic.
FILES
/usr/include/sys/ts.h
/etc/conf/pack.d/ts/space.c
SEE ALSO
priocntl(1), dispadmin(1M), idbuild(1M), mpcntl(1M), priocntl(2),
mpcntl(3X).
Page 4 Reliant UNIX 5.44 Printed 11/98