illumos: manual page: trapstat.8

TRAPSTAT(8) Maintenance Procedures TRAPSTAT(8)

NAME

trapstat - report trap statistics

SYNOPSIS

DESCRIPTION

The trapstat utility gathers and displays run-time trap statistics on
UltraSPARC-based systems. The default output is a table of trap types and
CPU IDs, with each row of the table denoting a trap type and each column
of the table denoting a CPU. If standard output is a terminal, the table
contains as many columns of data as can fit within the terminal width; if
standard output is not a terminal, the table contains at most six columns
of data. By default, data is gathered and displayed for all CPUs; if the
data cannot fit in a single table, it is printed across multiple tables.
The set of CPUs for which data is gathered and displayed can be
optionally specified with the -c or -C option.

Unless the -r option or the -a option is specified, the value displayed
in each entry of the table corresponds to the number of traps per second.
If the -r option is specified, the value corresponds to the number of
traps over the interval implied by the specified sampling rate; if the -a
option is specified, the value corresponds to the accumulated number of
traps since the invocation of trapstat.

By default, trapstat displays data once per second, and runs
indefinitely; both of these behaviors can be optionally controlled with
the interval and count parameters, respectively. The interval is
specified in seconds; the count indicates the number of intervals to be
executed before exiting. Alternatively, command can be specified, in
which case trapstat executes the provided command and continues to run
until the command exits. A positive integer is assumed to be an interval;
if the desired command cannot be distinguished from an integer, the full
path of command must be specified.

UltraSPARC I (obsolete), II, and III handle translation lookaside buffer
(TLB) misses by trapping to the operating system. TLB miss traps can be a
significant component of overall system performance for some workloads;
the -t option provides in-depth information on these traps. When run with
this option, trapstat displays both the rate of TLB miss traps and the
percentage of time spent processing those traps. Additionally, TLB misses
that hit in the translation storage buffer (TSB) are differentiated from
TLB misses that further miss in the TSB. (The TSB is a software structure
used as a translation entry cache to allow the TLB to be quickly filled;
it is discussed in detail in the UltraSPARC II User's Manual.) The TLB
and TSB miss information is further broken down into user- and kernel-
mode misses.

Workloads with working sets that exceed the TLB reach may spend a
significant amount of time missing in the TLB. To accommodate such
workloads, the operating system supports multiple page sizes: larger page
sizes increase the effective TLB reach and thereby reduce the number of
TLB misses. To provide insight into the relationship between page size
and TLB miss rate, trapstat optionally provides in-depth TLB miss
information broken down by page size using the -T option. The information
provided by the -T option is a superset of that provided by the -t
option; only one of -t and -T can be specified.

OPTIONS

The following options are supported:

-a
Displays the number of traps as accumulating,
monotonically increasing values instead of per-
second or per-interval rates.

-c cpulist
Enables trapstat only on the CPUs specified by
cpulist.

cpulist can be a single processor ID (for
example, 4), a range of processor IDs (for
example, 4-6), or a comma separated list of
processor IDs or processor ID ranges (for
example, 4,5,6 or 4,6-8).

-C processor_set_id
Enables trapstat only on the CPUs in the
processor set specified by processor_set_id.

trapstat modifies its output to always reflect
the CPUs in the specified processor set. If a CPU
is added to the set, trapstat modifies its output
to include the added CPU; if a CPU is removed
from the set, trapstat modifies its output to
exclude the removed CPU. At most one processor
set can be specified.

-e entrylist
Enables trapstat only for the trap table entry or
entries specified by entrylist. A trap table
entry can be specified by trap number or by trap
name (for example, the level-10 trap can be
specified as 74, 0x4A, 0x4a, or level-10).

entrylist can be a single trap table entry or a
comma separated list of trap table entries. If
the specified trap table entry is not valid,
trapstat prints a table of all valid trap table
entries and values. A list of valid trap table
entries is also found in The SPARC Architecture
Manual, Version 9 and the Sun Microelectronics
UltraSPARC II User's Manual. If the parsable
option (-P) is specified in addition to the -e
option, the format of the data is as follows:

Field Contents
1 Timestamp (nanoseconds since start)
2 CPU ID
3 Trap number (in hexadecimal)
4 Trap name
5 Trap rate per interval

Each field is separated with whitespace. If the
format is modified, it will be modified by adding
potentially new fields beginning with field 6;
exant fields will remain unchanged.

-l
Lists trap table entries. By default, a table is
displayed containing all valid trap numbers,
their names and a brief description. The trap
name is used in both the default output and in
the entrylist parameter for the -e argument. If
the parsable option (-P) is specified in addition
to the -l option, the format of the data is as
follows:

Field Contents
1 Trap number in hexadecimal
2 Trap number in decimal
3 Trap name
Remaining Trap description

-P
Generates parsable output. When run without other
data gathering modifying options (that is, -e, -t
or -T), trapstat's the parsable output has the
following format:

Field Contents
1 Timestamp (nanoseconds since start)
2 CPU ID
3 Trap number (in hexadecimal)
4 Trap name
5 Trap rate per interval

Each field is separated with whitespace. If the
format is modified, it will be modified by adding
potentially new fields beginning with field 6;
extant fields will remain unchanged.

-r rate
Explicitly sets the sampling rate to be rate
samples per second. If this option is specified,
trapstat's output changes from a traps-per-second
to traps-per-sampling-interval.

-t
Enables TLB statistics.

A table is displayed with four principal columns
of data: itlb-miss, itsb-miss, dtlb-miss, and
dtsb-miss. The columns contain both the rate of
the corresponding event and the percentage of CPU
time spent processing the event. The percentage
of CPU time is given only in terms of a single
CPU. The rows of the table correspond to CPUs,
with each CPU consuming two rows: one row for
user-mode events (denoted with u) and one row for
kernel-mode events (denoted with k). For each
row, the percentage of CPU time is totalled and
displayed in the rightmost column. The CPUs are
delineated with a solid line. If the parsable
option (-P) is specified in addition to the -t
option, the format of the data is as follows:

Field Contents
1 Timestamp (nanoseconds since start)
2 CPU ID
3 Mode (k denotes kernel, u denotes user)
4 I-TLB misses
5 Percentage of time in I-TLB miss handler
6 I-TSB misses
7 Percentage of time in I-TSB miss handler
8 D-TLB misses
9 Percentage of time in D-TLB miss handler
10 D-TSB misses
11 Percentage of time in D-TSB miss handler

Each field is separated with whitespace. If the
format is modified, it will be modified by adding
potentially new fields beginning with field 12;
extant fields will remain unchanged.

-T
Enables TLB statistics, with page size
information. As with the -t option, a table is
displayed with four principal columns of data:
itlb-miss, itsb-miss, dtlb-miss, and dtsb-miss.
The columns contain both the absolute number of
the corresponding event, and the percentage of
CPU time spent processing the event. The
percentage of CPU time is given only in terms of
a single CPU. The rows of the table correspond to
CPUs, with each CPU consuming two sets of rows:
one set for user-level events (denoted with u)
and one set for kernel-level events (denoted with
k). Each set, in turn, contains as many rows as
there are page sizes supported (see
getpagesizes(3C)). For each row, the percentage
of CPU time is totalled and displayed in the
right-most column. The two sets are delineated
with a dashed line; CPUs are delineated with a
solid line. If the parsable option (-P) is
specified in addition to the -T option, the
format of the data is as follows:

Field Contents
1 Timestamp (nanoseconds since start)
2 CPU ID
3 Mode k denotes kernel, u denotes user)
4 Page size, in decimal
5 I-TLB misses
6 Percentage of time in I-TLB miss handler
7 I-TSB misses
8 Percentage of time in I-TSB miss handler
9 D-TLB misses
10 Percentage of time in D-TLB miss handler
11 D-TSB misses
12 Percentage of time in D-TSB miss handler

Each field is separated with whitespace. If the
format is modified, it will be modified by adding
potentially new fields beginning with field 13;
extant fields will remain unchanged.

EXAMPLES

Example 1: Using trapstat Without Options

Example 2: Using trapset with CPU Filtering

The -c option can be used to limit the CPUs on which trapstat is
enabled. This example limits CPU 1 and CPUs 12 through 15.

example# trapstat -c 1,12-15

vct name | cpu1 cpu12 cpu13 cpu14 cpu15
------------------------+---------------------------------------------
24 cleanwin | 6923 3072 2500 3518 2261
44 level-4 | 3 0 0 1 1
49 level-9 | 100 100 100 100 100
4d level-13 | 23 8 14 19 14
60 int-vec | 2559 2699 2752 2688 2792
64 itlb-miss | 3296 1548 1174 1698 1087
68 dtlb-miss | 114788 54313 43040 58336 38057
6c dtlb-prot | 1046 549 417 545 370
84 spill-user-32 | 66551 29480 301588 26522 213032
88 spill-user-64 | 0 318652 111239 299829 221716
8c spill-user-32-cln | 856 347 331 416 293
90 spill-user-64-cln | 0 55 21 59 39
98 spill-kern-64 | 66464 31803 24758 34004 22277
a4 spill-asuser-32 | 1423 569 560 698 483
a8 spill-asuser-64 | 0 74 32 98 46
ac spill-asuser-32-cln | 4875 2250 1728 2384 1584
b0 spill-asuser-64-cln | 0 2 0 1 0
c4 fill-user-32 | 64193 28418 287516 27055 202093
c8 fill-user-64 | 0 305016 106692 288542 210654
cc fill-user-32-cln | 6733 3520 15185 2396 12035
d0 fill-user-64-cln | 0 13226 3506 12933 11032
d8 fill-kern-64 | 66220 31680 24674 33892 22196
108 syscall-32 | 2446 967 817 1196 755

Example 3: Using trapstat with TLB Statistics

The -t option displays in-depth TLB statistics, including the amount of
time spent performing TLB miss processing. The following example shows
that the machine is spending 14.1 percent of its time just handling D-TLB
misses:

example# trapstat -t
cpu m| itlb-miss %tim itsb-miss %tim | dtlb-miss %tim dtsb-miss %tim |%tim
-----+-------------------------------+-------------------------------+----
0 u| 2571 0.3 0 0.0 | 10802 1.3 0 0.0 | 1.6
0 k| 0 0.0 0 0.0 | 106420 13.4 184 0.1 |13.6
-----+-------------------------------+-------------------------------+----
1 u| 3069 0.3 0 0.0 | 10983 1.2 100 0.0 | 1.6
1 k| 27 0.0 0 0.0 | 106974 12.6 19 0.0 |12.7
-----+-------------------------------+-------------------------------+----
2 u| 3033 0.3 0 0.0 | 11045 1.2 105 0.0 | 1.6
2 k| 43 0.0 0 0.0 | 107842 12.7 108 0.0 |12.8
-----+-------------------------------+-------------------------------+----
3 u| 2924 0.3 0 0.0 | 10380 1.2 121 0.0 | 1.6
3 k| 54 0.0 0 0.0 | 102682 12.2 16 0.0 |12.2
-----+-------------------------------+-------------------------------+----
4 u| 3064 0.3 0 0.0 | 10832 1.2 120 0.0 | 1.6
4 k| 31 0.0 0 0.0 | 107977 13.0 236 0.1 |13.1
=====+===============================+===============================+====
ttl | 14816 0.3 0 0.0 | 585937 14.1 1009 0.0 |14.5

Example 4: Using trapstat with TLB Statistics and Page Size Information

By specifying the -T option, trapstat shows TLB misses broken down by
page size. In this example, CPU 0 is spending 7.9 percent of its time
handling user-mode TLB misses on 8K pages, and another 2.3 percent of its
time handling user-mode TLB misses on 64K pages.

example# trapstat -T -c 0
cpu m size| itlb-miss %tim itsb-miss %tim | dtlb-miss %tim dtsb-miss %tim |%tim
----------+-------------------------------+-------------------------------+----
0 u 8k| 1300 0.1 15 0.0 | 104897 7.9 90 0.0 | 8.0
0 u 64k| 0 0.0 0 0.0 | 29935 2.3 7 0.0 | 2.3
0 u 512k| 0 0.0 0 0.0 | 3569 0.2 2 0.0 | 0.2
0 u 4m| 0 0.0 0 0.0 | 233 0.0 2 0.0 | 0.0
- - - - - + - - - - - - - - - - - - - - - + - - - - - - - - - - - - - - - + - -
0 k 8k| 13 0.0 0 0.0 | 71733 6.5 110 0.0 | 6.5
0 k 64k| 0 0.0 0 0.0 | 0 0.0 0 0.0 | 0.0
0 k 512k| 0 0.0 0 0.0 | 0 0.0 206 0.1 | 0.1
0 k 4m| 0 0.0 0 0.0 | 0 0.0 0 0.0 | 0.0
==========+===============================+===============================+====
ttl | 1313 0.1 15 0.0 | 210367 17.1 417 0.2 |17.5

Example 5: Using trapstat with Entry Filtering

By specifying the -e option, trapstat displays statistics for only
specific trap types. Using this option minimizes the probe effect when
seeking specific data. This example yields statistics for only the dtlb-
prot and syscall-32 traps on CPUs 12 through 15:

example# trapstat -e dtlb-prot,syscall-32 -c 12-15
vct name | cpu12 cpu13 cpu14 cpu15
------------------------+------------------------------------
6c dtlb-prot | 817 754 1018 560
108 syscall-32 | 1426 1647 2186 1142

vct name | cpu12 cpu13 cpu14 cpu15
------------------------+------------------------------------
6c dtlb-prot | 1085 996 800 707
108 syscall-32 | 2578 2167 1638 1452

Example 6: Using trapstat with a Higher Sampling Rate

The following example uses the -r option to specify a sampling rate of
1000 samples per second, and filter only for the level-10 trap.
Additionally, specifying the -P option yields parsable output.

Notice the timestamp difference between the level-10 events: 9,998,000
nanoseconds and 10,007,000 nanoseconds. These level-10 events correspond
to the system clock, which by default ticks at 100 hertz (that is, every
10,000,000 nanoseconds).

example# trapstat -e level-10 -P -r 1000
1070400 0 4a level-10 0
2048600 0 4a level-10 0
3030400 0 4a level-10 1
4035800 0 4a level-10 0
5027200 0 4a level-10 0
6027200 0 4a level-10 0
7027400 0 4a level-10 0
8028200 0 4a level-10 0
9026400 0 4a level-10 0
10029600 0 4a level-10 0
11028600 0 4a level-10 0
12024000 0 4a level-10 0
13028400 0 4a level-10 1
14031200 0 4a level-10 0
15027200 0 4a level-10 0
16027600 0 4a level-10 0
17025000 0 4a level-10 0
18026000 0 4a level-10 0
19027800 0 4a level-10 0
20025600 0 4a level-10 0
21025200 0 4a level-10 0
22025000 0 4a level-10 0
23035400 0 4a level-10 1
24027400 0 4a level-10 0
25026000 0 4a level-10 0
26027000 0 4a level-10 0

ATTRIBUTES

See attributes(7) for descriptions of the following attributes:

+--------------------------------------------+
| ATTRIBUTE TYPE ATTRIBUTE VALUE |
|Interface Stability |
| Human Readable Output Unstable |
| Parsable Output Evolving |
+--------------------------------------------+

NOTES

When enabled, trapstat induces a varying probe effect, depending on the
type of information collected. While the precise probe effect depends
upon the specifics of the hardware, the following table can be used as a
rough guide:

Option Approximate probe effect
default 3-5% per trap
-e 3-5% per specified trap
-t, -T 40-45% per TLB miss trap
hitting in the TSB, 25-30%
per TLB miss trap missing
in the TSB

These probe effects are per trap not for the system as a whole. For
example, running trapstat with the default options on a system that
spends 7% of total time handling traps induces a performance degradation
of less than one half of one percent; running trapstat with the -t or -T
option on a system spending 5% of total time processing TLB misses induce
a performance degradation of no more than 2.5%.

When run with the -t or -T option, trapstat accounts for its probe effect
when calculating the %tim fields. This assures that the %tim fields are a
reasonably accurate indicator of the time a given workload is spending
handling TLB misses -- regardless of the perturbing presence of trapstat.

While the %tim fields include the explicit cost of executing the TLB miss
handler, they do not include the implicit costs of TLB miss traps (for
example, pipeline effects, cache pollution, etc). These implicit costs
become more significant as the trap rate grows; if high %tim values are
reported (greater than 50%), you can accurately infer that much of the
balance of time is being spent on the implicit costs of the TLB miss
traps.

Due to the potential system wide degradation induced, only the super-user
can run trapstat.

Due to the limitation of the underlying statistics gathering methodology,
only one instance of trapstat can run at a time.

April 9, 2016 TRAPSTAT(8)