System Tap Timestamp Notes
Revision as of 03:05, 27 October 2011 by Adushistova (talk | contribs)
This information was copied from the System Tap file tapsets/timestamp/timeset_tapset.txt, in May, 2006.
* Application name: sequence numbers and timestamps * Contact: Martin Hunt hunt@redhat.com Will Cohen wcohen@redhat.com Charles Spirakis charles.spirakis@intel.com * Motivation: On multi-processor systems, it is important to have a way to correlate information gathered between cpu's. There are two forms of correlation: a) putting information into the correct sequence order b) providing accurate time deltas between information If the resolution of the time deltas is high enough, it can also be used to order information. * Background: Discussion started due to relayfs and per-cpu buffers, but this is neede by many people. * Target software: Any software which wants to correlate data that was gathered on a multi-processor system, but the scope will be defined specifically for systemtap's needs. * Type of description: General information and discussion regarding sequencing and timing. * Interesting probe points: Any probe points where you are trying to get the time between two probe points. For example, timing how long a function takes and putting probe points at the function entry and function exit. * Interesting values: Possible ways to order data from multiple sources include: Retrieve the sequence/time from a global area High Precision Event Timer (HPET) Possible implementation: multimedia/HPET timer arch/i386/kernel/timer_hpet.c Advantages: granularity can vary (HPET spec says minimum freq of HPET timer is 10Mhz =~100ns resolution), can be treated as read-only, can bypass cache update and avoid being caced at all if desired, desigend to be used as an smp timestamp (see specification) Disadvantages: may not be available on all platforms, may not be synchronized on NUMA systems (ie counts for all processors within a numa node is comparable, but counts for processors between nodes may not be comparable), potential resource conflict if timers used by other software Real Time Clock (RTC) Possible implementation: "external" chip (clock chip) which has time information, accessed via ioport or memory-mapped io Advantages: can be treated as read-only, can bypass cache update and avoid being cached at all if desired Disadvantages: may not be available on all platforms, low granularity (for rtc, ~1ms), usually slow access ACPI Power Management Timer (pm timer) Possible implementation: implemented as part of the ACPI specification at 3.579545Mhz arch/i386/kernel/timers/timer_pm.c Advantages: not affected by throttling, halting or power saving states, moderate granularity (3.5Mhz ~ 300ns resolution), desigend for use by an OS to keep track of time during sleep/power states Disadvantages: may not be available on all platforms, slower access than hpet timer (but still much faster than RTC) Chipset counter Possible implementation: timer on a processor chipset, ??SGI implementation??, do we know of any other implementations ? Advantages: likely to be based on pci bus clock (33Mhz = ~30ns) or front-side-bus clock (200Mhz = ~5ns) Disadvantages: may not be available on all platforms Sequence Number Possible implementation: atomic_t global variable, cache aligned, placed in struct to keep variable on a cache line by itself Advantages: guaranteed correct ordering (even on NUMA systems), architecure independent, platform independent Disadvantages: potential for cache-line ping-pong, doesn't scale, no time information (ordering data only), access can be slower on NUMA systems Jiffies Possible implementation: OS counts the number of "clock interrupts" since power on. Advantages: platform independent, architecture independent, one writer, many readers (less cache ping-pong) cached at all if desired Disadvantages: low resolution (usually 10ms, sometimes 1ms). Do_gettimeofday Possible implementation: arch/i386/kernel/time.c Advantages: platform independent, architecture independent, 1 writer, many readers (less cache ping-pong), accuracy of micro-seconds Disadvantages: the time unit increment value used by this routine changes based on information from ntp (i.e if ntp needs to speed up / slow down the clock, then callers to this routine will be affected). This is a disadvantage for timing short intervals, but an advantage for timing long intervals. Retrieve the sequence/time from a cpu-unique area Timestamp counter (TSC) Possible implementation: count of the number of core cycles the processor has executed since power on, due to lack of synchronization between cpus, would also need to keep track of which cpu the tsc came from Advantages: no external bus access, high granularity (cpu core cycles), available on most (not all) architectures, platform independent Disadvantages: not synchronized between cpus, since it is a count of cpu cycles count can be affected by throttling, halting and power saving states, may not correlate to "actual" time (ie, just because a 1G processor showed a delta of 1G cycles, doesn't mean 1 second has passed) APIC timer Possible implementation: timer implemented within the processor Advantages: no external bus access, moderate to high granularity (usually counting based front-side bus clock or core clock) Disadvantages: not synchronized between cpus, may be affected by throttling, halting/power saving states, may not correlate to "actual" time. PMU event Possible implementation: program a perfmonance counter with a specific event related to time Advantages: no external bus access, moderate to high granularity (usually counting based front-side bus clock or core clock), can be virtualized to give moderate to high granularity for individual thread paths Disadvantages: not synchronized between cpus, may be affected by throttling, halting/power saving states, may not correlate to "actual" time, processor dependent For reference, as a quick baseline, on Martin's dual-processor system, he gets the following performance measurements: kprobe overhead: 1200-1500ns (depending on OS and processor) atomic read plus increment: 40ns (single processor access, no conflict) monotonic_clock() 550ns do_gettimeofday() 590ns * Dependencies: Not Applicable * Restrictions: Certain timers may already be in use by other parts of the kernel depending on how it is configured (for example, RTC is used by the watchdog code). Some kernels may not compile in the necessary code (for example, if using the pm timer, need ACPI). Some platforms or architectures may not have the timer requested (for example, there is no HPET timer on older systesm). * Data collection: For data collection, it is probably best to keep the concept between sequence ordering and timestamp separate within systemtap (for both the user as well as the implementation). For sequence ording, the initial implementation should use ?? the atomic_t form for the sequence ordering (since it is guaranteed to be platform and architecture neutral)?? and modify/change the implementation later if there is a problem. For timestamp, the initial implementation should use ?? hpet timer ?? pm timer ?? do_gettimeofday ?? cpu # + tsc ?? some combination (do_gettimeofday + cpu # & low bits of tsc)? We could do something like what LTT does (see below) to generate 64-bit timestamps containing the nanoseconds since Jan 1, 1970. Assuming the implementation keeps these concepts separate now (ordering data vs. timing deltas), it is always possible to merge them in the future if a high granularity, numa/smp synchronized timesource becomes available for a large number of platforms and/or processors. * Data presentation: In general, users prefer output which is based on "actual" time (ie, they prefer an output that says the delta is XXX nanoseconds instead of YYY cpu cycles). Most of the time users want delta's (how long did this take), but occasionally they want absolute times (when / what time was this information collected) * Competition: DTRACE has output in nanoseconds (and it is comparable between processors on an mp system), but it is unclear what the actual resolution is. Even if the sparc machine does have hardware that provides nanosecond resolution, on x86-64 they are likely to have the same problems as discussed here since the solaris opteron box tends to be a pretty vanilla box. From Joshua Stone (joshua.i.stone at intel.com): == BEGIN === DTrace gives you three built-in variables: uint64_t timestamp: The current value of a nanosecond timestamp counter. This counter increments from an arbitrary point in the past and should only be used for relative computations. uint64_t vtimestamp: The current value of a nanosecond timestamp counter that is virtualized to the amount of time that the current thread has been running on a CPU, minus the time spent in DTrace predicates and actions. This counter increments from an arbitrary point in the past and should only be used for relative time computations. uint64_t walltimestamp: The current number of nanoseconds since 00:00 Universal Coordinated Time, January 1, 1970. As for how they are implemented, the only detail I found is that timestamp is "similar to the Solaris library routine gethrtime". The manpage for gethrtime is here: http://docs.sun.com/app/docs/doc/816-5168/6mbb3hr8u?a=view == END == What LTT does: "Cycle counters are fast to read but may reflect time inaccurately. Indeed, the exact clock frequency varies with time as the processor temperature changes, influenced by the external temperature and its workload. Moreover, in SMP systems, the clock of individual processors may vary independently. LTT corrects the clock inaccuracy by reading the real time clock value and the 64 bits cycle counter periodically, at the beginning of each block, and at each 10ms. This way, it is sufficient to read only the lower 32 bits of the cycle counter at each event. The associated real time value may then be obtained by linear interpolation between the nearest full cycle counter and real time values. Therefore, for the average cost of reading and storing the lower 32 bits of the cycle counter at each event, the real time with full resolution is obtained at analysis time." * Cross-references: The profile_tapset is very dependent on sequencing and time when ordering of data (i.e. taking a trace history) as well as high granularity (when calculating time deltas). * Associated files: Profile tapset requirements .../src/tapsets/profile/profile_tapset.txt Intel high precision event timers specification: http://www.intel.com/hardwaredesign/hpetspec.htm ACPI specification: http://www.acpi.info/DOWNLOADS/ACPIspec-2-0b.pdf From an internal email sent by Tony Luck (tony.luck at intel.com) regarding a clustered environment. For the summary below, hpet and pm timer were not an option. For systemtap, they should be considered, especially since pm timer and hpet were designed to be a timestamp. == BEGIN === For extremely short intervals (<100ns) get some h/w help (oscilloscope or logic analyser). Delays reading TSC and pipeline effects could skew your results horribly. Having a 2GHz clock doesn't mean that you can measure 500ps intervals. For short intervals (100ns to 10 ms) TSC is your best choice ... but you need to sample it on the same cpu, and converting the difference between two TSC values to real time will require some system dependent math to find the nominal frequency of the system (you may be able to ignore temperature effects, unless your system is in an extremely hostile environment). But beware of systems that change the TSC rate when making frequency adjustments for power saving. It shouldn't be hard to measure the system clock frequency to about five significant digits of accuracy, /proc/cpuinfo is probably good enough. For medium intervals (10 ms to a minute) then "gettimeofday()" or "clock_gettime()" on a system *NOT* running NTP may be best, but you will need to adjust for systematic error to account for the system clock running fast/slow. Many Linux systems ship with a utility named "clockdiff" that you can use to measure the system drift against a reference system (a system that is nearby on the network, running NTP, prefereably a low "stratum" one). Just run clockdiff every five minutes for an hour or two, and plot the results to see what systematic drift your system has without NTP. N.B. if you find the drift is > 10 seconds per day, then NTP may have trouble keeping this system synced using only drift corrections, you might see "steps" when running NTP. Check /var/log/messages for complaints from NTP. For long intervals (above a minute). Then you need "gettimeofday()" on a system that uses NTP to keep it in touch with reality. Assuming reasonable network connectivity, NTP will maintain the time within a small number of milliseconds of reality ... so your results should be good for 4-5 significant figures for 1 minute intervals, and better for longer intervals. == END ==