[an error occurred while processing this directive]

HP OpenVMS Systems Documentation
Content starts here

OpenVMS Alpha System Analysis Tools Manual

Previous Contents Index

Chapter 6
SDA Spinlock Tracing Utility

This chapter presents an overview of the SDA Spinlock Tracing Utility commands, and describes the SDA Spinlock Tracing commands.

6.1 Overview of the SDA Spinlock Tracing Utility

To synchronize access to data structures, the OpenVMS operating system uses a set of static and dynamic spinlocks, such as IOLOCK8 and SCHED. The operating system acquires a spinlock to synchronize data, and at the end of the critical code path the spinlock is then released. If a CPU attempts to acquire a spinlock while another CPU is holding it, the CPU attempting to acquire the spinlock has to spin, waiting until the spinlock is released. Any lost CPU cycles within such a spinwait loop are charged as MPsynch time.

By using the MONITOR utility, you can monitor the time in process modes, for example, with the command $ MONITOR MODES. A high rate of MP synchronization indicates contention for spinlocks. However, until the implementation of the Spinlock Tracing utility, there was no way to tell which spinlock was heavily used, and who was acquiring and releasing the contended spinlocks. The Spinlock Tracing utility allows a characterization of spinlock usage. It can also collect performance data for a given spinlock on a per-CPU basis.

This tracing ability is built into the system synchronization execlet, which contains the spinlock code, and can be enabled or disabled while the system is running. There is no need to reboot the system to load a separate debug image. The images that provide spinlock tracing functionality are as follows:

SYS$LOADABLE_IMAGES:SPL$DEBUG.EXE
SYS$SHARE:SPL$SDA.EXE

The SDA> prompt provides the command interface. From this command interface, you can load and unload the spinlock debug execlet using SPL LOAD and SPL UNLOAD, and start, stop and display spinlock trace data. This allows you to collect spinlock data for a given period of time without system interruption. Once information is collected, the trace buffer can be deallocated and the execlet can be unloaded to free up system resources. The spinlock trace buffer is allocated from S2 space and pages are taken from the freelist.

Should the system crash while spinlock tracing is enabled, the trace buffer is dumped into the system dump file, and it can later be analyzed using the spinlock trace utility. This is very useful in tracking down CPUSPINWAIT bugcheck problems.

Note that by enabling spinlock tracing, there is a performance impact. The amount of the impact depends on the amount of spinlock usage.

Note
The Spinlock Tracing utility is still under development. The command format, displays, and suggested approach to spinlock analysis are all subject to change.

6.2 How to Use the SDA Spinlock Tracing Utility

The following steps will enable you to collect spinlock statistics using the Spinlock Tracing Utility.

  1. Load the Spinlock Tracing Utility execlet. 


    SDA> SPL LOAD
  2. Allocate a trace buffer and start tracing. 


    SDA> SPL START TRACE
  3. Wait a few seconds to allow some tracing to be done, then find out which spinlocks are incurring the most acquisitions and the most spinwaits. 


    SDA> SPL SHOW TRACE/SUMMARY

    For example, you might see contention for the SCHED and IOLOCK8 spinlocks (a high acquisition count, with a significant proportion of the acquisitions being forced to wait).
  4. Look to see if the spinlocks with a high proportion of spinwaits caused a significant delay in the acquisition of the spinlock. You must now collect more detailed statistics on a specific spinlock. 


    SDA> SPL START COLLECT/SPINLOCK=SCHED

    This command accumulates additional data for the specified spinlock. As long as tracing is not stopped, collection will continue to accumulate spinlock-specific data from the trace buffer.
  5. Display the additional data collected for the specified spinlock. 


    SDA> SPL SHOW COLLECT

    This display includes the average hold time of the spinlock and the average spinwait time while acquiring the spinlock.
  6. Repeat steps 4 and 5 for each spinlock that has contention. A START COLLECT cancels the previous collection.
  7. Disable spinlock tracing when you have collected all the needed spinlock statistics and release all the memory used by the Spinlock Tracing utility with the following commands. 


    SDA> SPL STOP COLLECT
SDA> SPL STOP TRACE
SDA> SPL UNLOAD

6.3 Example Command Procedure for Collection of Spinlock Statistics

The following example shows a command procedure that can be used for gathering spinlock statistics: 


$ analyze/system
  spl load
  spl start trace/buffer=1000
  spawn wait 00:00:15
  spl stop trace
  read/executive/nolog
  set output spl_trace.lis
  spl show trace/summary
  spl start collect/spin=sched
  spawn wait 00:00:05
  spl show collect
  spl start collect/spin=iolock8
  spawn wait 00:00:05
  spl show collect
  spl start collect/spin=lckmgr
  spawn wait 00:00:05
  spl show collect
  spl start collect/spin=mmg
  spawn wait 00:00:05
  spl show collect
  spl start collect/spin=timer
  spawn wait 00:00:05
  spl show collect
  spl start collect/spin=mailbox
  spawn wait 00:00:05
  spl show collect
  spl start collect/spin=perfmon
  spawn wait 00:00:05
  spl show collect
  spl stop collect
  spl unload
  exit
$ exit

A more comprehensive procedure is provided as SYS$EXAMPLES:SPL.COM.

6.4 Listing of SDA Spinlock Tracing Commands

The following is a list of the spinlock tracing commands:

SPL LOAD
SPL SHOW COLLECT
SPL SHOW TRACE
SPL START COLLECT
SPL START TRACE
SPL STOP COLLECT
SPL STOP TRACE
SPL UNLOAD

SPL LOAD

Loads the SPL$DEBUG execlet. This must be done prior to starting spinlock tracing.

Format

SPL LOAD

Parameters

None.

Qualifiers

None.

Description

The SPL LOAD command loads the SPL$DEBUG execlet, which contains the tracing routines.

Example


SDA> SPL LOAD
SPL$DEBUG load status = 00000001

SPL SHOW COLLECT

Displays the collected spinlock data.

Format

SPL SHOW COLLECT [/RATES|/TOTALS]

Parameters

None.

Qualifiers

/RATES

Reports activity as a rate per second and hold/spin time as a percentage of time. This is the default.

/TOTALS

Reports activity as a count and hold/spin time as cycles.

Description

The SPL SHOW COLLECT command displays the collected spinlock data. It displays first a summary on a per-CPU basis, followed by the callers of the specific spinlock. This second list is sorted by the top consumers of the spinlock (in percent of time held). These displays show average spinlock hold and spinlock wait time in system cycles.
Example

SPL SHOW TRACE

Displays spinlock tracing information.

Format

SPL SHOW TRACE [/[NO]SPINLOCK=spinlock|/[NO]FORKLOCK=forklock
|/[NO]ACQUIRE|/RATES |/[NO]RELEASE|/[NO]WAIT
|/[NO]FRKDSPTH|/[NO]FRKEND
|/SUMMARY|/CPU=n |/TOP=n|/TOTALS]

Parameters

None.

Qualifiers

/SPINLOCK=spinlock
/NOSPINLOCK

The /SPINLOCK=n qualifier specifies the display of a specific spinlock, for example, /SPINLOCK=LCKMGR or /SPINLOCK=SCHED.

The /NOSPINLOCK qualifier specifies that no spinlock trace information be displayed. If omitted, all spinlock trace entries are decoded and displayed.

/FORKLOCK=forklock
/NOFORKLOCK

The /FORKLOCK=forklock qualifier specifies the display of a specific forklock, for example, /FORKLOCK=IOLOCK8 or /FORKLOCK=IPL8.

The /NOFORKLOCK qualifier specifies that no forklock trace information be displayed. If omitted, all fork trace entries are decoded and displayed.

/ACQUIRE
/NOACQUIRE

The /ACQUIRE qualifier displays any spinlock acquisitions.

The /NOACQUIRE qualifier ignores any spinlock acquisitions.

/RATES

Reports activity as a rate per second and hold/spin time as a percentage of time. This is the default.

/RELEASE
/NORELEASE

The /RELEASE qualifier displays any spinlock releases.

The /NORELEASE qualifier ignores any spinlock releases.

/TOTALS

Reports activity as a count and hold/spin time as cycles.

/WAIT
/NOWAIT

The /WAIT qualifier displays any spinwait operations.

The /NOWAIT qualifier ignores any spinwait operations.

/FRKDSPTH
/NOFRKDSPTH

The /FRKDSPTH qualifier displays all invocations of fork routines within the fork dispatcher. This is the default.

The /NOFRKDSPTH qualifier ignores all of the operations of the /FRKDSPTH qualifier.

/FRKEND
/NOFRKEND

The /FRKEND qualifier displays all returns from fork routines within the fork dispatcher. This is the default.

The /NOFRKEND qualifier ignores all operations of the /FRKEND qualifier.

/CPU=n

Specifies the display of information for a specific CPU only, for example, /CPU=5 or /CPU=PRIMARY. By default, all trace entries for all CPUs are displayed.

/SUMMARY

Steps through the entire trace buffer and displays a summary of all spinlock and forklock activity. It also displays the top ten callers.

/TOP=n

Displays a different number other than the top ten callers or fork PCs. By default, the top ten are displayed. This qualifier is only useful when you also specify the /SUMMARY qualifier.

Description

The SPL SHOW TRACE command displays spinlock tracing information. The latest acquired or released spinlock is displayed first, and then the trace buffer is stepped backwards in time.

By default, all trace entries will be displayed, but you can use qualifiers to select only certain entries.

Since this is not a time critical activity and a table lookup has to be done anyway to translate the SPL address to a spinlock name, commands like /SPINLOCK=(SCHED,IOLOCK8) do work. /SUMMARY will step the entire trace buffer and display a summary of all spinlock activity, along with the top-ten callers' PCs. You can use /TOP=n to display a different number of the top ranked callers.

Examples

Callout Meaning
1 Shows timestamps that are collected as system cycle counters (SCC) and then displayed with an accuracy down to microseconds. Each CPU is incrementing its own SCC as soon as it is started, so there is some difference between different CPUs' system cycle counters. The standard system time is incremented only every 10 Msec and as such is not exact enough. Adjusting the SCC to the specific CPU's system time and translating it into an accurate timestamp will thus sometimes display times out of order for different CPUs. However, for the same CPU ID, the timestamps are accurate.
2 Shows the physical CPU ID of the CPU logging the trace entry.
3 Shows the address of the spinlock fork. If it is a static one, its name is displayed; otherwise, it is marked as ???.
4 Shows the caller's PC address that acquired or released the spinlock, or the fork PC if the trace entry is a forklock. Symbolization is attempted, so a READ/EXECUTIVE might help to display a routine name, instead of simply a module and offset.
5 Shows the EPID, which is the external PID of the process generating the trace entry. If an interrupt or fork was responsible for the entry, then a zero EPID is displayed.
6 Shows the trace operation. For a spinlock, which was acquired without going through a spinwait, there is a matching acquire/release pair of trace entries for the same CPU ID for a given spinlock. If a spinlock is held, it cannot be acquired immediately, so there is also a spinwait trace entry for this pair. The different variations of the acquire and release operations are distinguished, as are the same spinlocks if they are acquired recursively multiple times.
7 Shows the address of the trace buffer entry, in case there is a need to access the raw and undecoded trace data.
Callout Meaning
8 Shows the summary information by stepping through the whole trace buffer, and displaying a single line of information for each spinlock. If the percent of spin wait is very high, then a spinlock is a candidate for high contention.
9 For each spinlock in the summary display, the top ten callers' PCs are displayed along with the number of spinlock acquisitions and releases, as well as spinwait counts and the number of multiple acquisitions of the same spinlock.
Callout Meaning
10 The forklock summary displays the number of fork operations on a specific CPU for each forklock. For each forklock, the top ten fork PC addresses are displayed, along with the minimum, maximum and average duration of the fork operation in system cycles. The percent of time spent in a given fork routine is displayed along with the percent of time for the forklock.

SPL START COLLECT

Starts to collect spinlock information a longer period of time than will fit into the trace buffer.

Format

SPL START COLLECT [/SPINLOCK=spinlock|/ADDRESS=n]

Parameters

None.

Qualifiers

/SPINLOCK=spinlock

Specifies the tracing of a specific spinlock, for example, /SPINLOCK=LCKMGR or /SPINLOCK=SCHED.

/ADDRESS=n

Specifies the tracing of a specific spinlock by address.

Description

The SPL START COLLECT command starts a collection of spinlock information for a longer period of time than will fit into the trace buffer. You need to enable spinlock tracing before a spinlock collection can be started. On a system with heavy activity, the trace buffer typically can only hold a relatively small time window of spinlock information. In order to collect spinlock information over a longer time period, a collection can be started. The collection tries to catch up with the running trace index and save the spinlock information into a balanced tree within the virtual address space of the process performing the spinlock collection. Either use the name of a static spinlock, or supply the address of a dynamic spinlock, for which information should be gathered.

The trace entries are kept in the trace buffer, which is allocated from S2 space, hence there is no disruption, if tracing is started from within SDA and then the user exits from SDA. However, for the longer period data collection, the information is kept in process-specific memory, thus a user needs to stay within SDA; otherwise the data collection is automatically terminated by SDA's image rundown. You can collect data for two or more spinlocks simultaneously, by using a separate process for each collection.

Examples

#1 

SDA> SPL START COLLECT
Use /SPINLOCK=name or /ADDRESS=n to specify which spinlock info needs to be collected...

This example shows that you need to supply either a spinlock name of a static spinlock, or the address of a dynamic spinlock, if you want to collect information over a long period of time.

#2 

SDA> SPL START COLLECT/SPINLOCK=LCKMGR

This example shows the command line to start to collect information on the usage of the LCKMGR spinlock.

SPL START TRACE

Enables spinlock tracing.

Format

SPL START TRACE [/[NO]SPINLOCK=spinlock|/[NO]FORKLOCK=forklock
|/BUFFER=pages|/[NO]ACQUIRE|
|/[NO]RELEASE|/[NO]WAIT|/[NO]FRKDSPTH
|/[NO]FRKEND|/CPU=n]

Parameters

None.

Qualifiers

/SPINLOCK=spinlock
/NOSPINLOCK

The /SPINLOCK=spinlock qualifier specifies the tracing of a specific spinlock, for example, /SPINLOCK=LCKMGR or /SPINLOCK=SCHED.

The /NOSPINLOCK qualifier disables spinlock tracing and does not collect any spinlock data. If omitted, all spinlocks are traced.

/FORKLOCK=forklock
/NOFORKLOCK

The /FORKLOCK=forklock qualifier specifies the tracing of a specific forklock, for example, /FORKLOCK=IOLOCK8 or /FORKLOCK=IPL8.

The /NOFORKLOCK qualifier disables forklock tracing and does not collect any forklock data. If omitted, all forks are traced.

/BUFFER=pages

Specifies the size of the trace buffer (in Alpha page units). It defaults to 128 pages, which is equivalent to 1MB, if omitted.

/ACQUIRE
/NOACQUIRE

The /ACQUIRE qualifier traces any spinlock acquisitions. This is the default.

The /NOACQUIRE qualifier ignores any spinlock acquisitions.

/RELEASE
/NORELEASE

The /RELEASE qualifier traces any spinlock releases. This is the default.

The /NORELEASE qualifier ignores any spinlock releases.

/WAIT
/NOWAIT

The /WAIT qualifier traces any spinwait operations. This is the default.

The /NOWAIT qualifier ignores any spinwait operations.

/FRKDSPTH
/NOFRKDSPTH

The /FRKDSPTH qualifier traces all invocations of fork routines within the fork dispatcher. This is the default.

The /NOFRKDSPTH qualifier ignores all of the /FRKDSPTH operations.

/FRKEND
/NOFRKEND

The /FRKEND qualifier traces all returns from fork routines within the fork dispatcher. This is the default.

The /NOFRKEND qualifier ignores all of the operations of the /FRKEND qualifier.

/CPU=n

Specifies the tracing of a specific CPU only, for example, /CPU=5 or /CPU=PRIMARY. By default, all CPUs are traced.

Description

The SPL START TRACE command enables spinlock and fork tracing. By default all spinlocks and forks are traced and a 128 page (1MByte) trace buffer is allocated and used as a ring buffer.

Examples

#1 

SDA> SPL START TRACE/BUFFER=1000
Tracing started... (Spinlock = 00000000, Forklock = 00000000)

This example shows how to enable a tracing for all spinlock and forklock operations into a 8 MByte trace buffer.

#2 

SDA> SPL START TRACE/CPU=PRIMARY/SPINLOCK=SCHED /NOFORKLOCK
Tracing started... (Spinlock = 810AF600, Forklock = 00000000)

This example shows how to trace only SCHED spinlock operations on the primary CPU.

#3 

SDA> SPL START TRACE /NOSPINLOCK /FORKLOCK=IPL8
Tracing started... (Spinlock = 00000000, Forklock = 863A4C00)

This example shows how to trace only fork operations to IPL8.

Previous Next Contents Index