1 \documentclass{article}
2 \usepackage{epsfig,placeins}
5 % Copyright (C) 2007-2009 Alan D. Brunelle <Alan.Brunelle@hp.com>
7 % This program is free software; you can redistribute it and/or modify
8 % it under the terms of the GNU General Public License as published by
9 % the Free Software Foundation; either version 2 of the License, or
10 % (at your option) any later version.
12 % This program is distributed in the hope that it will be useful,
13 % but WITHOUT ANY WARRANTY; without even the implied warranty of
14 % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 % GNU General Public License for more details.
17 % You should have received a copy of the GNU General Public License
18 % along with this program; if not, write to the Free Software
19 % Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 % vi :set textwidth=75
23 \title{\texttt{btt} User Guide}
24 \author{Alan D. Brunelle (Alan.Brunelle@hp.com)}
30 \section{\label{sec:intro}Introduction}
32 \texttt{btt} is a post-processing tool for the block layer IO tracing
33 tool called blktrace. As noted in its Users Guide, blktrace
36 is a block layer IO tracing mechanism which provides detailed
37 information about request queue operations up to user space.
40 blktrace is capable of producing tremendous amounts of output in the
41 form of multiple individual traces per IO executed during the traced
42 run. It is also capable of producing some general statistics concerning
43 IO rates and the like. \texttt{btt} goes further and produces a variety
44 of overall statistics about each of the individual handling of IOs, and
45 provides data we believe is useful to plot to provide visual comparisons
48 This document will discuss \texttt{btt} usage, provide some sample output,
49 and also show some interesting plots generated from the data provided
50 by the \texttt{btt} utility.
53 A short note on the ordering of this document -- the actual
54 command-line usage section occurs relatively late in the document (see
55 section~\ref{sec:cmd-line}), as we felt that discussing some of the
56 capabilities and output formats would make the parameter discussion
60 This document refers to the output formats generated by \texttt{btt}
61 version 2.00. However, the descriptions are general enough to cover
62 output formats prior to that.
64 \newpage\tableofcontents
66 \newpage\section{\label{sec:getting-started}Getting Started}
68 The simple pipeline to get going with \texttt{btt} is to perform the
72 \item Run \texttt{blktrace}, specifying whatever devices and other
73 parameters you want. You must save the traces to disk in this step,
74 btt does not work in live mode.
76 \item After tracing completes, run \texttt{blkrawverify}, specifying
77 all devices that were traced (or at least on all devices that you
78 will use \texttt{btt} with -- section~\ref{sec:o-D} shows how you
79 can dictate which devices to use with btt). If blkrawverify finds
80 errors in the trace streams saved, it is best to recapture the data
81 -- utilizing \texttt{btt} on \emph{unclean} trace files produces
84 While this step is optional, we have found that performing this
85 helps to ensure data coming from \texttt{btt} makes the most sense.
87 \item Run \texttt{blkparse} with the \texttt{-d} option specifying
88 a file to store the combined binary stream. (e.g.: \texttt{blkparse
91 \texttt{blktrace} produces a series of binary files
92 containing parallel trace streams -- one file per CPU per
93 device. \texttt{blkparse} provides the ability to combine all the
94 files into one time-ordered stream of traces for all devices.
96 \item Run \texttt{btt} specifying the file produced by
97 \texttt{blkparse} utilizing the \texttt{-i} option (e.g.: \texttt{btt
102 \newpage\section{\label{sec:output-overview}Output Overview}
104 The major default areas of output provided by \texttt{btt}
105 include\label{tl-defs}:
108 \item[average component times across all IOs] The time line of each IO
109 is broken down into 3 major regions:
112 \item Time needed to insert or merge an incoming IO onto the request
113 queue. This is the average time from when the IO enters the block
114 IO layer (queue trace) until it is inserted (insert trace).
116 This is denoted as \emph{Q2I} time.
118 This is also broken down into two component times\footnote{On
119 occasion there are also some time spent \emph{sleeping} waiting
120 for a request. That occurs between the Q and G operations. You
121 will see these listed as \texttt{S2G} times.}:
124 \item[Q2G] Time needed to \emph{get} a request (get request
127 \item[G2I] Time needed to put that request onto the request
128 queue (insert trace).
131 For \emph{merged} requests -- an incoming request that is merged
132 with a previously submitted request -- we calculate \emph{Q2M}, the
133 amount of time between the queue trace and the merge trace.
135 \item Time spent on the request queue. The average time from when
136 the IO is inserted or merged onto the request queue, until it is
137 issued (issue trace) to the lower level driver.
139 Referred to as \emph{I2D} time\footnote{The \emph{issue} trace
140 is represented by a D in the blkparse output, hence its usage in
141 btt to refer to issue traces. Note that an I is used to refer to
142 \emph{insert} traces.}.
144 \item Driver and device time -- the average time from when the
145 actual IO was issued to the driver until is completed (completion
146 trace) back to the block IO layer.
148 This is referred to as the \emph{D2C} time\
151 Two other sets of results are presented in this section:
154 \item \emph{Q2Q} which measures the time between queue traces
155 in the system. This provides some idea as to how quickly IOs are
156 being handed to the block IO layer.
158 \item \emph{Q2C} which measures the times for the complete life cycle
159 of IOs during the run\footnote{One of the areas that needs some
160 work in \texttt{btt} is to better understand the multiplex nature of
161 IOs during a run. In theory, one would like ${Q2I} + {I2D} + {D2C}
162 = {Q2C}$ however, typically there are multiple queue traces that
163 are combined via merges into a single IO issued and completed. We
164 currently average the queue-to-insert and queue-to-merge times,
165 and thus tend to be quite close to the expected equation.}
169 For each row in this output, we provide a minimum, average, maximum
170 (which are all presented in seconds), and overall count. As an
171 example\footnote{As with this display, the author has taken some liberty
172 in reformatting the output for better display on the printed page.}:
176 ---- ------------- ------------- ------------- -----------
177 Q2Q 0.000000058 0.000012761 9.547941661 2262310
178 Q2I 0.000000272 0.000005995 0.104588839 2262311
179 I2D 0.000001446 0.094992714 0.239636864 2262311
180 D2C 0.000193721 0.030406554 1.634221408 2262311
181 Q2C 0.000207665 0.125405263 1.830917198 2262311
184 When tracking \emph{device mapper} devices, we also break down the
185 \emph{Q2A} and \emph{Q2C} times for those IOs.
187 \item[Device Overhead]
189 Using the data from the previous chart, we can then provide some idea
190 as to where IO spend most of the time on average. The following output
191 shows the percentage of time spent in each of the phases of an
192 IO\footnote{It should be noted that incoming requests either go through:
201 before proceeding to I2D and D2C.}
204 DEV | Q2G G2I Q2M I2D D2C
205 ---------- | --------- --------- --------- --------- ---------
206 ( 8, 80) | 0.0013% 0.0004% 0.0006% 88.5005% 11.4988%
207 ---------- | --------- --------- --------- --------- ---------
208 Overall | 0.0003% 0.0001% 0.0041% 21.4998% 78.4958%
211 \item[Device Merge Information]
213 A key measurement when making changes in the system (software \emph{or}
214 hardware) is to understand the block IO layer ends up merging incoming
215 requests into fewer, but larger, IOs to the underlying driver. In this
216 section, we show the number of incoming requests (Q), the number of
217 issued requests (D) and the resultant ratio. We also provide values
218 for the minimum, average and maximum IOs generated.
220 Looking at the following example:
223 DEV | #Q #D Ratio | BLKmin BLKavg BLKmax Total
224 ---------- | ------- ----- ----- | ------ ------ ------ -------
225 ( 68, 64) | 2262311 18178 124.5 | 2 124 128 2262382
228 we see that (on average) the block IO layer is combining upwards of
229 125 incoming requests into a single request down the IO stack. The
230 resultant average IO size is 124 blocks.
232 \item[Device Seek Information]
234 Another useful measure is the variability in the sector distances
235 between consecutively \emph{received -- queued} and \emph{submitted
236 -- issued} IOs. The next two sections provides some rudimentary
237 statistics to gauge the general nature of the sector differences
238 between IOs. Values provided include the number of seeks (number of IOs
239 submitted to lower level drivers), the \emph{mean} distance between
240 IOs, the \emph{median} value for all seeks, and the \emph{mode} -
241 the value(s) and the counts are provided for the latter.
243 The first of the two sections displays values for Q2Q seek distances --
244 providing a set of indicators showing how close incoming IO requests
245 are to each other. The second section shows D2D seek distances --
246 providing a set of indicators showing how close the IO requests are
247 that are handled by underlying drivers.
250 DEV | NSEEKS MEAN MEDIAN | MODE
251 --------- | ------ ------- ------ | -------
252 ( 68, 64) | 18178 19611.3 0 | 0(17522)
255 We have almost exclusively seen median and mode values of 0, indicating
256 that seeks tend to have an equal amount of forward and backwards
257 seeks. The larger the count for the mode in comparison to the total
258 number of seeks is indicative as to how many IOs are coming out of
259 the block IO layer in adjacent sectors. (Obviously, the higher this
260 percentage, the better the underlying subsystems can handle them.)
262 \item[Request Queue Plug Information]
264 During normal operation, requests queues are \emph{plugged} and during
265 such times the IO request queue elements are not able to be processed
266 by underlying drivers. The next section shows how often the request
267 queue was in such a state.
270 DEV | # Plugs # Timer Us | % Time Q Plugged
271 --------- | ------- ---------- | ----------------
272 ( 68, 64) | 833( 0) | 0.356511895%
275 There are two major reasons why request queues are unplugged, and both
276 are represented in the above table.
279 \item Explicit unplug request from some subsystem in the kernel.
281 \item Timed unplugs, due to a request queue exceeding some temporal
282 limit for being plugged.
285 The total number of unplugs is equal to the number of plugs less the
286 ones due to timer unplugs.
288 \item[IOs per Unplug \& Unplugs-due-to-timeout]
290 In this subsection one can see the average number of IOs on the request
291 queue at the time of an unplug or unplug due to a timeout. The following
292 sample shows a sample of both unplug sections:
295 ==================== Plug Information ====================
297 DEV | # Plugs # Timer Us | % Time Q Plugged
298 ---------- | ---------- ---------- | ----------------
299 ( 8, 0) | 1171( 123) | 0.280946640%
300 ( 8, 32) | 4( 0) | 0.000325469%
301 ---------- | ---------- ---------- | ----------------
302 Overall | # Plugs # Timer Us | % Time Q Plugged
303 Average | 587( 61) | 0.140636055%
305 DEV | IOs/Unp IOs/Unp(to)
306 ---------- | ---------- ----------
309 ---------- | ---------- ----------
310 DEV | IOs/Unp IOs/Unp(to)
314 This table and the preceding one have to be considered together --
315 in the sample output in the immediately preceding table one can see
316 how the larger number of data values for device (8,0) dominates in
319 \newpage\item[Active Requests At Q Information]
321 An important consideration when analyzing block IO schedulers is to
322 know how many requests the scheduler has to work with. The metric
323 provided in this section details how many requests (on average) were
324 being held by the IO scheduler when an incoming IO request was being
325 handled. To determine this, \texttt{btt} keeps track of how many Q
326 requests came in, and subtracts requests that have been issued (D).
328 Here is a sample output of this sections:
331 ==================== Active Requests At Q Information ====================
334 ---------- | -------------
339 ---------- | -------------
340 Overall | Avgs Reqs @ Q
344 \item[I/O Active Period Information]
346 In this subsection data is tabulated showing I/O activity on a
347 per-device as well across all devices being traced. ``I/O activity''
348 is defined as periods of time when the underlying device driver and
349 device have at least one I/O to work upon. The values presented include:
352 \item[\# Live] Number of periods of ``liveness.''
353 \item[Avg. Act] Average length of each period ov ``liveness.''
354 \item[Avg. !Act] Aerage length of each non-active period.
355 \item[\% Live] Percent of total time spent with the driver/device active.
358 Here is a sample portion of this type of chart:
361 DEV | # Live Avg. Act Avg. !Act % Live
362 ---------- | ---------- ------------- ------------- ------
363 ( 8, 16) | 29 0.909596815 0.094646263 90.87
364 ( 8, 32) | 168 0.097848226 0.068231948 59.06
365 ---------- | ---------- ------------- ------------- ------
366 Total Sys | 33 0.799808811 0.082334758 90.92
369 For information on generating data files that can be plotted with
370 per-device and system-wide I/O activity see section~\ref{sec:o-Z}.
375 \subsection*{\label{sec:detailed-data}Detailed Data}
377 In addition to the default sections output, if one supplies the
378 \texttt{--all-data} or \texttt{-A} argument (see section~\ref{sec:o-A})
379 to \texttt{btt} further sections are output:
382 \item[Per Process] As traces are emitted, they are tagged with the
383 process ID of the currently running thread in the kernel. The process
384 names are also preserved, and mapped to the ID. For each of the parts
385 of the time line discussed above on page~\pageref{tl-defs}, a chart is
386 provided which breaks down the traces according to process ID (name).
388 One must be aware, however, that the process ID may not have anything
389 to do with the originating IO. For example, if an application is
390 doing buffered IO, then the actual submitted IOs will most likely
391 come from some page buffer management daemon thread (like pdflush,
392 or kjournald for example). Similarly, completion traces are rarely
393 (if ever?) going to be associated with the process which submitted
394 the IO in the first place.
396 Here is a sample portion of this type of chart, showing Q2Q times
401 ------------- ----------- ----------- ----------- -------
402 mkfs.ext3 0.000000778 0.000009074 1.797176188 1899371
403 mount 0.000000885 0.000672513 0.030638128 73
404 pdflush 0.000000790 0.000006752 0.247231307 179791
407 \item[Per Process Averages] The average columns from the above charts,
408 are also presented in their own chart.
410 \item[Per Device] Similar to the per-process display, \texttt{btt}
411 will also break down the various parts of an IOs time line based upon a
412 per-device criteria. Here's a portion of this area, displayed showing
413 the issued to complete times (D2C).
417 --------- ----------- ----------- ----------- ------
418 ( 65, 80) 0.000140488 0.001076906 0.149739869 169112
419 ( 65, 96) 0.000142762 0.001215221 0.173263182 155488
420 ( 65,112) 0.000145221 0.001254966 0.124929936 165726
421 ( 65,128) 0.000141896 0.001159596 0.775231052 169015
422 ( 65,144) 0.000140832 0.001290985 0.211384698 210661
423 ( 65,160) 0.000139915 0.001175554 0.073512063 133973
424 ( 65,176) 0.000141254 0.001104870 0.073231310 145764
425 ( 65,192) 0.000141453 0.001234460 0.167622507 140618
429 \item[Per Device Averages] The average columns from the above charts,
430 are also presented in their own chart.
432 \item[Q2D Histogram] A display of histogram buckets for the Q to D times
433 -- basically, from where an IO enters the block IO layer for a given
434 device, and when it is dispatched. The buckets are arranged via the
435 time in seconds, as in:
438 ==================== Q2D Histogram ====================
440 DEV | <.005 <.010 <.025 <.050 <.075 <.100 <.250 <.500 < 1.0 >=1.0
441 --------- | ===== ===== ===== ===== ===== ===== ===== ===== ===== =====
442 ( 66, 80) | 61.2 7.9 12.1 7.9 3.0 1.4 1.5 0.2 0.0 4.6
443 ( 65,192) | 42.3 5.0 8.7 30.0 8.9 3.0 1.8 0.1 0.0 0.1
444 ( 65,128) | 34.3 5.3 8.9 32.0 9.7 3.7 5.3 0.6 0.0 0.1
446 ( 65, 64) | 59.9 4.2 6.0 24.6 4.2 0.8 0.1 0.0 0.0 0.1
447 ( 66, 64) | 62.6 8.1 12.7 7.9 2.4 0.6 0.1 0.0 0.0 5.4
448 ========== | ===== ===== ===== ===== ===== ===== ===== ===== ===== =====
449 AVG | 52.9 6.2 10.0 20.1 5.3 1.7 1.4 0.2 0.0 2.1
454 \newpage\section{\label{sec:data-files}Data Files Output}
456 Besides the averages output by default, the following 5(+) files are also
457 created with data points which may be plotted.
460 \item[\emph{file}.dat] This file provides a notion of \emph{activity}
461 for the system, devices and processes. The details of this file are
462 provided in section~\ref{sec:activity}.
464 \item[\emph{file}\_qhist.dat] Provides histogram data for the size of
465 incoming IO requests, for more information see section~\ref{sec:qhist}.
467 \item[\emph{file}\_dhist.dat] Provides histogram data for the size
468 of IO requests submitted to lower layer drivers, for more information
469 see section~\ref{sec:dhist}.
471 \item[\emph{file}\_mbps\_fp.dat] Provides a set of data for
472 mb-per-second values each second
473 - for more information see section~\ref{sec:rstat}.
475 \item[\emph{file}\_iops\_fp.dat] Provides a set of data for
476 I/Os-per-second values each second
477 - for more information see section~\ref{sec:rstat}.
481 In addition to the default data files output, there are optional data
482 files which can be generated by btt. These include:
485 \item[subset of \texttt{.avg} data, easily parsed ] When the
486 \texttt{-X} option is specified \emph{and} the \texttt{-o} has also
487 been specified, then a subset of the data produced by default is
488 copied to another file that is \emph{more easily parsed.} Refer to
489 section~\ref{sec:o-X} for full details.
491 \item[iostat] iostat-like data can be distilled by btt, and is
492 described in section~\ref{sec:iostat}.
494 \item[per IO detail] Each and every IO traced can be output in a form
495 that shows each of the IO components on consecutive lines (rather
496 than grepping through a blkparse output file for example). The
497 details on this file is included in section~\ref{sec:per-io}.
499 \item[iostat] Latency information -- both Q2d, D2c and Q2C --
500 on a per-IO basis can be generated. These are described in
501 section~\ref{sec:lat}.
503 \item[seek details] A set of data files containing all IO-to-IO
504 sector differences can be output, with details found in
505 section~\ref{sec:seek}.
507 \item[unplug histogram details] A data file per device containing
508 histogram output for the amount of IOs released at unplug time.
509 Section~\ref{sec:o-u} has more details.
512 \newpage\section{\label{sec:activity}Activity Data File}
514 The activity data file contains a series of data values that indicate
515 those periods of time when queue and complete traces are being
516 processed. The values happen to be in a format easily handled by
517 xmgrace\footnote{\texttt{http://plasma-gate.weizmann.ac.il/Grace/}
518 ``Grace is a WYSIWYG 2D plotting tool for the X Window System and
519 M*tif.''}, but is easy to parse for other plotting and/or analysis
522 The file is split into pairs of sets of data points, where each pair
523 contains a set of queue activity and a set of completion activity. The
524 points are presented with the first column (X values) being the time
525 (in seconds), and the second column (Y values) providing an on/off
526 type of setting. For each pair, the Y values have two settings off
527 (low) and on (high). For example, here is a snippet of a file showing
532 # Total System : q activity
555 What this indicates is that there was q activity for the system
556 from 0.000000000 through 0.000070381, but was inactive from there to
557 1.023482637, and so on. Section~\ref{sec:o-d} contains details on how
558 to adjust btt's notion of what constitutes activity.
560 The pairs are arranged as follows:
563 \item First there is the total system activity -- meaning activity
564 in either queue or completion traces across all devices.
566 \item Next comes per-device activity information -- for each device
567 being traced, that request queues Q and C traces are presented.
569 \item Last we present pairs per-process.
572 Using this, one is then able to plot regions of activity versus
573 inactivity -- and one can gather a sense of deltas between the queueing
574 of IOs and when they are completed. Figure~\ref{fig:activity} shows
575 a very simplistic chart showing some activity:
578 \leavevmode\centering
579 \epsfig{file=activity.eps,width=4.5in}
580 \caption{\label{fig:activity}Simple Activity Chart}
583 When the black line (system Q activity) is \emph{high}, then the system
584 is seeing relatively continuous incoming queues. Conversely, when it is
585 low, it represents an extended period of time where no queue requests
586 were coming in. Similarly for the red line and C activity.
588 \newpage\section{\label{sec:hist}Histogram Data Files}
590 The histogram data files provide information concerning incoming and
591 outgoing IO sizes (in blocks). For simplicity, the histogram buckets
592 are one-for-one for sizes up to 1,024 blocks in the IO, and then a
593 single bucket for all sizes greater than or equal to 1,024 blocks.
595 The files are again in grace-friendly format, with the first set
596 containing data for the first 1,023 buckets, and a separate set
597 representing sizes $\ge 1024$ blocks. (This is done so that one can
598 easily use a separate formatting specification for the latter set.)
600 The first column (X values) is the various IO sizes, and the second
601 column (Y values) represents the number of IOs of that size.
603 \subsection*{\label{sec:qhist}Q Histogram Data File}
605 Figure~\ref{fig:qhist} is a sample graph generated from data used during
606 some real-world analysis\footnote{Note the logarithmic nature of the
607 Y axis for this chart.}. With the visual representation provided by
608 this, one can quickly discern some different characteristics between
609 the 3 runs -- in particular, one can see that there is only a single
610 red point (representing 8 blocks per IO), whereas the other two had
611 multiple data points greater than 8 blocks.
614 \leavevmode\centering
615 \epsfig{file=qhist.eps,width=4.5in}
616 \caption{\label{fig:qhist}Q Histogram}
619 \subsection*{\label{sec:dhist}D Histogram Data File}
621 Figure~\ref{fig:dhist} is a sample graph generated from data used during
622 some real-world analysis\footnote{Note the logarithmic nature of the
623 Y axis for this chart.}. Again, visually, one can see that the black
624 and blue dots are somewhat similar below about 192 blocks per IO going
625 out. And then one can make the broad generalization of higher reds,
626 lower blues and blacks in the middle.
629 \leavevmode\centering
630 \epsfig{file=dhist.eps,width=4.5in}
631 \caption{\label{fig:dhist}D Histogram}
634 \newpage\section{\label{sec:rstat}Running Stats Files}
636 There are two files produced for each of all devices being traced
637 (prefixed with \emph{sys\_}) and per-device (prefixed with the device
640 The two files are for reporting I/O rate (I/Os per second - name ends
641 with \texttt{iops\_fp.dat}) and throughput (MiB per second - name ends
642 with \texttt{mbps\_fp.dat}).
644 The data in the files has two columns:\smallskip
647 \textbf{File Type} & \textbf{X values} & \textbf{Y values}\\\hline
648 \textbf{iops} & Runtime (seconds) & I/Os per second\\\hline
649 \textbf{mbps} & Runtime (seconds) & MiB per second\\\hline
662 These can be plotted using various tools (e.g., xmgrace as in
663 figure~\ref{fig:rstats}).
666 \leavevmode\centering
667 \epsfig{file=rstats.eps,width=4.5in}
668 \caption{\label{fig:rstats}Running Stats}
671 \newpage\section{\label{sec:iostat}iostat Data File}
672 \texttt{btt} attempts to produce the results from running an
673 \texttt{iostat -x} command in parallel with the system as it is being
674 traced. The fields (columns) generated by the \texttt{--iostat} or
675 \texttt{-I} option can be seen from the following output snippet --
676 note that the line has been split to fit on the printed page:
679 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s
680 rkB/s wkB/s avgrq-sz avgqu-sz await svctm %util Stamp
682 ( 8, 16) 0.00 0.00 0.00 1005.30 0.00 152806.36
683 0.00 76403.18 152.00 31.00 0.00 0.00 0.00 71.79
685 ( 8, 16) 1.02 5.80 0.34 1.07 4.03 55.62
686 2.02 27.81 42.13 0.61 0.00 21.90 0.00 TOTAL
689 Note that the STAMP field contains the runtime (in seconds) for that
692 \newpage\section{\label{sec:per-io}Per-IO Data File}
694 \texttt{btt} can produce a text file containing time line data for each
695 IO processed. The time line data contains rudimentary information for
696 the following stages:
700 \item get request traces
704 \item completion traces
708 The \emph{--per-io-dump} or \emph{-p} option triggers this behavior,
709 and will produce a file containing streams of IOs (separated by blank
710 spaces). As an example, here is a snippet of 4 IOs that were merged
711 together, you will note there are 3 merged IOs, and 1 inserted in the
712 stream. The issue and completion traces are replicated per IO.
715 66,0 : 0.763283556 Q 6208+8
718 0.763338848 D 6208+32
719 0.763705760 C 6208+32
721 66,0 : 0.763314550 Q 6224+8
723 0.763338848 D 6208+32
724 0.763705760 C 6208+32
726 66,0 : 0.763321010 Q 6232+8
728 0.763338848 D 6208+32
729 0.763705760 C 6208+32
731 65,240: 0.763244173 Q 6216+8
733 0.763374288 D 6208+32
734 0.763826610 C 6208+32
737 The columns provide the following information:
740 \item Device major/minor.
742 \item Time of the trace (seconds from the start of the run)
746 \item start block + number of blocks
749 \newpage\section{\label{sec:lat}\label{sec:lat-q2d}\label{sec:lat-q2c}\label{sec:lat-d2c}Latency Data Files}
751 The latency data files which can be optionally produced by \texttt{btt}
752 provide per-IO latency information, one for queue time (Q2D), one
753 for total IO time (Q2C) and one for latencies induced by lower layer
754 drivers and devices (D2C).
756 In both cases, the first column (X values) represent runtime (seconds),
757 while the second column (Y values) shows the actual latency for a
758 command at that time (either Q2D, D2C or Q2C).
760 \newpage\section{\label{sec:seek}Seek Data Files}
762 \texttt{btt} can also produce two data files containing all IO-to-IO sector
763 deltas, providing seek information which can then be plotted. The
764 produced data file contains 3 sets of data:
767 \item Combined data -- all read and write IOs
769 \item Read data -- just seek deltas for reads
771 \item Write data -- just seek deltas for writes
774 The format of the output file names is to have the name generated by
775 the following fields separated by underscores (\texttt{\_}):
778 \item The prefix provided as the argument to the \texttt{-s} option.
779 \item The major and minor numbers of the device separated by a comma.
780 \item The string \texttt{q2q} or \texttt{d2d}, indicating the Q2Q or
781 D2D seeks, respectively.
782 \item One of the following characters:
784 \item[r] For read (device to system) IOs
785 \item[w] For write (system to device) IOs
786 \item[c] Combined -- both read and write IOs
790 An example name would be after specifying \texttt{-s seek} would be:
791 \texttt{seek\_065,048\_q2q\_w.dat}.
793 The format of the data is to have the runtime values (seconds since
794 the start of the run) in column 1 (X values); and the difference in
795 sectors from the previous IO in column 2 (Y values). Here is a snippet
796 of the first few items from a file:
800 0.000034733 35283790.0
801 0.000106453 35283790.0
802 0.005239009 35283950.0
803 0.006968575 35283886.0
804 0.007218709 35283694.0
805 0.012145393 35283566.0
806 0.014980835 -35848914.0
807 0.024239323 -35848914.0
808 0.024249402 -35848914.0
809 0.025707095 -35849072.0
813 Figure~\ref{fig:seek} shows a simple graph that can be produced which
814 provides visual details concerning seek patterns.
817 \leavevmode\centering
818 \epsfig{file=seek.eps,width=4.5in}
819 \caption{\label{fig:seek}Seek Chart}
823 The seek difference is calculated in one of two ways:
826 \item[default] By default, the seek distance is calculated as the
827 \emph{closest} distance between the previous IO and this IO. The
828 concept of \emph{closeness} means that it could either be the
829 \emph{end} of the previous IO and the beginning of the next, or the
830 end of this IO and the start of the next.
832 \item[\texttt{-a}] If the \texttt{-a} or \texttt{--seek-absolute}
833 option is specified, then the seek distance is simply the difference
834 between the end of the previous IO and the start of this IO.
837 \newpage\subsection{\label{sec:sps-spec}Seeks Per Second}
839 When the \texttt{-m} option provides a name, Q2Q and/or D2D seeks
840 will trigger \texttt{btt} to output seeks-per-second information. The
841 first column will contain a time value (seconds), and the second column
842 will indicate the number of seeks per second at that point.
844 When there is only a single data point within a 1-second window,
845 \texttt{btt} will just output the time value for the point, and the
846 value 1.0 in the second column. If there is no perceived difference
847 in the times present for the current sample, then the second columns
848 value is the number of seeks present at that time.
850 Otherwise, if $\alpha$ and $\Omega$ are the first and last times
851 seen within a 1-second window, and $\nu$ are the number of seeks seen
852 in that time frame, then:
855 \item[column 1] Midway point in time for this span, or: \hfill$\alpha +
856 {{(\Omega - \alpha)} / 2}$
858 \item[column 2] Average seeks per second over this span, or: \hfill$\nu /
862 Figure~\ref{fig:sps} shows a simple pair of graphs generated from
866 \leavevmode\centering
867 \epsfig{file=sps.eps,width=4.5in}
868 \caption{\label{fig:sps}Seeks-per-second Chart}
872 \newpage\section{\label{sec:cmd-line}Command Line}
876 [ -a | --seek-absolute ]
878 [ -B <output name> | --dump-blocknos=<output name> ]
879 [ -d <seconds> | --range-delta=<seconds> ]
880 [ -D <dev;...> | --devices=<dev;...> ]
881 [ -e <exe,...> | --exes=<exe,...> ]
883 [ -i <input name> | --input-file=<input name> ]
884 [ -I <output name> | --iostat=<output name> ]
885 [ -l <output name> | --d2c-latencies=<output name> ]
886 [ -L <freq> | --periodic-latencies=<freq> ]
887 [ -m <output name> | --seeks-per-second=<output name> ]
888 [ -M <dev map> | --dev-maps=<dev map>
889 [ -o <output name> | --output-file=<output name> ]
890 [ -p <output name> | --per-io-dump=<output name> ]
891 [ -P <output name> | --per-io-trees=<output name> ]
892 [ -q <output name> | --q2c-latencies=<output name> ]
893 [ -Q <output name> | --active-queue-depth=<output name> ]
895 [ -s <output name> | --seeks=<output name> ]
896 [ -S <interval> | --iostat-interval=<interval> ]
897 [ -t <sec> | --time-start=<sec> ]
898 [ -T <sec> | --time-end=<sec> ]
899 [ -u <output name> | --unplug-hist=<output name> ]
902 [ -X | --easy-parse-avgs ]
903 [ -z <output name> | --q2d-latencies=<output name> ]
907 \subsection{\label{sec:o-a}\texttt{--seek-absolute}/\texttt{-a}}
909 When specified on the command line, this directs btt to calculate
910 seek distances based solely upon the ending block address of one IO,
911 and the start of the next. By default \texttt{btt} uses the concept
912 of the closeness to either the beginning or end of the previous IO. See
913 section~\ref{sec:seek} for more details about seek distances.
915 \subsection{\label{sec:o-A}\texttt{--all-data}/\texttt{-A}}
917 Normally \texttt{btt} will not print out verbose information
918 concerning per-process and per-device data (as outlined in
919 section~\ref{sec:detailed-data}). If you desire that level of
920 detail you can specify this option.
922 \subsection{\label{sec:o-B}\texttt{--dump-blocknos}/\texttt{-B}}
924 This option will output absolute block numbers to three files prefixed
925 by the specified output name:
928 \item[\emph{prefix}\_\emph{device}\_r.dat] All read block numbers are
929 output, first column is time (seconds), second is the block number,
930 and the third column is the ending block number.
932 \item[\emph{prefix}\_\emph{device}\_w.dat] All write block numbers are
933 output, first column is time (seconds), second is the block number,
934 and the third column is the ending block number.
936 \item[\emph{prefix}\_\emph{device}\_c.dat] All block numbers (read
937 and write) are output, first column is time (seconds), second is
938 the block number, and the third column is the ending block number.
941 \subsection{\label{sec:o-d}\texttt{--range-delta}/\texttt{-d}}
943 Section~\ref{sec:activity} discussed how \texttt{btt} outputs a file
944 containing Q and C activity, the notion of \emph{active} traces simply
945 means that there are Q or C traces occurring within a certain period
946 of each other. The default values is 0.1 seconds; with this option
947 allowing one to change that granularity. The smaller the value, the
948 more data points provided.
950 \subsection{\label{sec:o-D}\texttt{--devices}/\texttt{-D}}
952 Normally, \texttt{btt} will produce data for all devices detected in
953 the traces parsed. With this option, one can reduce the analysis to
954 one or more devices provided in the string passed to this option. The
955 device identifiers are the major and minor number of each device, and
956 each device identifier is separated by a colon (:). A valid specifier
957 for devices 8,0 and 8,8 would then be: \texttt{"8,0:8,8"}.
959 \subsection{\label{sec:o-e}\texttt{--exes}/\texttt{-e}}
961 Likewise, \texttt{btt} will produce data for all processes (executables)
962 found in the traces. With this option, one can specify which processes
963 you want displayed in the output. The format of the string passed is
964 a list of executable \emph{names} separated by commas (,). An example
965 would be \texttt{"-e mkfs.ext3,mount"}.
967 \subsection{\label{sec:o-h}\texttt{--help}/\texttt{-h}}
969 Prints out the simple help information, as seen at the top of
970 section~\ref{sec:cmd-line}.
972 \subsection{\label{sec:o-i}\texttt{--input-file}/\texttt{-i}}
974 Specifies the binary input file that \texttt{btt} will interpret traces
975 in. See section~\ref{sec:getting-started} for information concerning
978 \subsection{\label{sec:o-I}\texttt{--iostat}/\texttt{-I}}
980 This option triggers \texttt{btt} to generate iostat-like output to the
981 file specified. Refer to section~\ref{sec:iostat} for more information
982 on the output produced.
984 \subsection{\label{sec:o-l}\texttt{--d2c-latencies}/\texttt{-l}}
986 This option instructs \texttt{btt} to generate the D2C latency file
987 discussed in section~\ref{sec:lat-d2c}.
989 \subsection{\label{sec:o-L}\texttt{--periodic-latencies}/\texttt{-L}}
991 When given a value greater than 0, this option will create two data
992 files (q2c \& d2c) per device containing a periodic timestamp \&
993 average latency over that period.
995 \subsection{\label{sec:o-m}\texttt{--seeks-per-second}\texttt{-m}}
997 Tells \texttt{btt} to output seeks per second information. Each device
998 being measured can have up to 2 files output: One with Q2Q information
999 and one with D2D seek information. Information on the output produced
1000 can be found in section~\ref{sec:sps-spec}.
1003 \textbf{Note: This requires seek output to be selected -- see
1004 section~\ref{sec:seek}.}
1007 \subsection{\label{sec:o-M}\texttt{--dev-maps}/\texttt{-M}}
1009 Internal option, still under construction.
1011 \subsection{\label{sec:o-o}\texttt{--output-file}/\texttt{-o}}
1013 Normally \texttt{btt} sends the statistical output (covered in
1014 section~\ref{sec:output-overview}) to standard out, if you specify
1015 this option this data is redirected to the file specified.
1017 \subsection{\label{sec:o-p}\texttt{--per-io-dump}/\texttt{-p}}
1019 This option tells \texttt{btt} to generate the per IO dump file as
1020 discussed in section~\ref{sec:per-io}.
1022 \subsection{\label{sec:o-P}\texttt{--per-io-tress}/\texttt{-P}}
1024 The \texttt{-P} option will generate a file that contains a list of all IO
1025 "sequences" - showing only the Q, D \& C operation times. The D \& C
1026 time values are separated from the Q time values with a vertical bar.
1028 \subsection{\label{sec:o-q}\texttt{--q2c-latencies}/\texttt{-q}}
1030 This option instructs \texttt{btt} to generate the Q2C latency file
1031 discussed in section~\ref{sec:lat-q2c}.
1033 \subsection{\label{sec:o-Q}\texttt{--active-queue-depth}/\texttt{-Q}}
1035 This option tells \texttt{btt} to generate a data file (using the given
1036 name as a base) which contains: A time stamp in the first column,
1037 and then the number of \emph{active} requests issued to the device
1038 driver. (The value is incremented when an \emph{issue} is performend,
1039 and decremented when a \emph{complete} is performed.
1041 \subsection{\label{sec:o-r}\texttt{--no-remaps}/\texttt{-r}}
1043 Ignore remap traces; older kernels did not implement the full remap PDU.
1045 \subsection{\label{sec:o-s}\texttt{--seeks}/\texttt{-s}}
1047 This option instructs \texttt{btt} to generate the seek data file
1048 discussed in section~\ref{sec:seek}.
1050 \subsection{\label{sec:o-S}\texttt{--iostat-interval}/\texttt{-S}}
1052 The normal \texttt{iostat} command allows one to specify the snapshot
1053 interval, likewise, \texttt{btt} allows one to specify how many seconds
1054 between its generation of snapshots of the data via this option. Details
1055 about the iostat-like capabilities of \texttt{btt} may be found in
1056 section~\ref{sec:iostat}.
1058 \subsection{\label{sec:o-tT}\texttt{--time-start}/\texttt{-t} and
1059 \texttt{--time-end}/\texttt{T}}
1062 \emph{This \texttt{btt} capability is still under construction, results are
1063 not always consistent at this point in time.}
1066 These options allow one to dictate to \texttt{btt} when to start and stop
1067 parsing of trace data in terms of seconds since the start of the run. The
1068 trace chosen will be between the start time (or 0.0 if not
1069 specified) and end time (or the end of the run) specified.
1071 \subsection{\label{sec:o-u}\texttt{--unplug-hist}/\texttt{-u}}
1073 This option instructs \texttt{btt} to generate a data file containing
1074 histogram information for \emph{unplug} traces on a per device
1075 basis. It shows how many times an unplug was hit with a specified
1076 number of IOs released. There are 21 output values into the file, as
1081 \textbf{X value} & \textbf{Representing Counts} \\\hline
1085 \dots & \dots\dots\\
1091 The file name(s) generated use the text string passed as an argument for
1092 the prefix, followed by the device identifier in \texttt{major,minor}
1093 form, with a \texttt{.dat} extension (as an example, with \texttt{-u
1094 up\_hist} specified on the command line: \texttt{up\_hist\_008,032.dat}.
1096 \subsection{\label{sec:o-V}\texttt{--version}/\texttt{-V}}
1098 Prints out the \texttt{btt} version, and exits.
1100 \subsection{\label{sec:o-v}\texttt{--verbose}/\texttt{-v}}
1102 While \texttt{btt} is processing data, it will put out periodic (1-second
1103 granularity) values describing the progress it is making through the
1104 input trace stream. The value describes how many traces have been
1105 processed. At the end of the run, the overall number of traces, trace
1106 rate (number of thousands of traces per second), and the real time for
1107 trace processing and output are displayed. Example (note: the interim
1108 trace counts are put out with carriage returns, hence, they overwrite
1112 # btt -i bp.bin -o btt -v
1113 Sending range data to bttX.dat
1114 Sending stats data to bttX.avg
1119 4581291 traces @ 279.7 Ktps
1120 16.379036+0.000005=16.379041
1123 \subsection{\label{sec:o-X}\texttt{--easy-parse-avgs}/\texttt{-X}}
1125 \emph{Some} of the data produced by default can also be shipped
1126 simultaneously to another file in an easy to parse form. When
1127 the \texttt{-o} option is selected (thus producing a file with a
1128 \texttt{.avg} exentsion), \emph{and} the \texttt{-X} flag is present,
1129 then \texttt{btt} will generate this file.
1131 The format is space-delimited values starting with a 3-character
1132 \emph{record} indicator, then the device information (either major,minor
1133 or the device name when \texttt{-M} is specified), and then a number of
1134 fields representing data values. The following table shows the record
1135 identifiers and the fields provided:
1138 \begin{tabular}{|l|l|}\hline
1139 \textbf{Record} & \textbf{Description}\\\hline
1140 \texttt{DMI} & Device Merge Information:\\
1141 & \#Q \#D Ratio BLKmin BLKavg BLKmax Total\\\hline
1142 \texttt{QSK} & Device Q2Q Seek Information:\\
1143 & NSEEKS MEAN MEDIAN MODE N-MODE mode\ldots\\\hline
1144 \texttt{DSK} & Device D2D Seek Information:\\
1145 & NSEEKS MEAN MEDIAN MODE N-MODE mode\ldots\\\hline
1146 \texttt{PLG} & Plug Information:\\
1147 & \#Plugs \#TimerUnplugs \%TimeQPlugged\\\hline
1148 \texttt{UPG} & Unplug Information:\\
1149 & IOsPerUnplug IOsPerUnplugTimeout\\\hline
1150 \texttt{ARQ} & Active Requests at Q Information:\\
1151 & AvgReqs@Q\\\hline\hline
1152 \texttt{Q2Q} & Queue-to-Queue times:\\
1153 \texttt{Q2G} & Queue-to-GetRequest times:\\
1154 \texttt{S2G} & Sleep-to-GetRequest times:\\
1155 \texttt{G2I} & GetRequest-to-Insert times:\\
1156 \texttt{Q2M} & Queue-to-Merge times:\\
1157 \texttt{I2D} & Insert-to-Issue times:\\
1158 \texttt{M2D} & Merge-to-Issue times:\\
1159 \texttt{D2C} & Issue-to-Complete times:\\
1160 \texttt{Q2C} & Queue-to-Complete times:\\
1161 & MIN AVG MAX N\\\hline
1165 A sample output file would look like:
1168 Q2Q 0.000000001 0.003511356 9.700000000 309906
1169 Q2G 0.000000001 0.774586535 805.300000000 106732
1170 S2G 0.000000001 0.072525952 0.370000000 578
1171 G2I 0.000000001 0.000001125 0.010000000 106732
1172 Q2M 0.000000001 0.730763626 751.820000000 204040
1173 I2D 0.000000001 1.270720538 612.880000000 106948
1174 M2D 0.000000001 0.992355230 428.930000000 203114
1175 D2C 0.000000001 0.008681311 137.020000000 307343
1176 Q2C 0.000000001 1.304370794 805.660000000 308921
1177 DMI 8,16 309907 106729 2.903681286 8 182 1024 19504768
1178 QSK 8,16 309907 167200.935561314 0 0 235708
1179 DSK 8,16 106729 433247.436563633 0 0 33974
1180 PLG 8,16 40824 382 0.008881420
1181 UPG 8,16 1.993361748 1.866492147
1182 ARQ 8,16 12.938165321
1185 \subsection{\label{sec:o-z}\texttt{--q2d-latencies}/\texttt{-z}}
1187 This option instructs \texttt{btt} to generate the Q2D latency file
1188 discussed in section~\ref{sec:lat-q2d}.
1190 \subsection{\label{sec:o-Z}\texttt{--q2d-latencies}/\texttt{-Z}}
1192 This option generates per-device (and total system) data files. Each
1193 file contain a data line which resembles a timing graph: low meaning
1194 I/O inactive, high meaning I/O active. A sample section of two ``active'' regions would be:
1208 Which shows an active area from 0.000025733 through
1209 0.000107089 followed by another at 0.005637386 through
1210 0.017323909. Figure~\ref{fig:live_plot} shows a sample plot that can
1211 be generated by such data.
1214 \leavevmode\centering
1215 \epsfig{file=live.eps,width=5.5in}
1216 \caption{\label{fig:live_plot}Sample graph using data from \texttt{-Z}}
1219 \newpage\section{\label{sec:bno_plot}bno\_plot.py}
1221 Included with the distribution is a simple 3D plotting utility based
1222 upon the block numbers output when \texttt{-B} is specified (see
1223 section~\ref{sec:o-B} for more details about the \texttt{-B option}). The
1224 display will display \emph{each} IO generated, with the time (seconds)
1225 along the X-axis, the block number (start) along the Y-axis and the
1226 number of blocks transferred in the IO represented along the Z-axis.
1228 The script requires Python\footnote{\texttt{www.python.org}} and
1229 gnuplot\footnote{\texttt{www.gnuplot.info}}, and will enter interactive
1230 mode after the image is produced. In this interactive mode one can enter
1231 gnuplot commands at the \texttt{'gnuplot>'} prompt, and/or can change
1232 the viewpoint within the 3D image by \emph{left-click-hold} and moving
1233 the mouse. A sample screen shot can be seen in figure~\ref{fig:bno_plot}.
1235 \subsection*{\texttt{bno\_plot.py} Command Line Options}
1240 $ bno_plot.py --help
1244 [ -K | --keys-below ]
1248 Utilizes gnuplot to generate a 3D plot of the block number
1249 output from btt. If no <files> are specified, it will
1250 utilize all files generated after btt was run with -B
1251 blknos (meaning: all files of the form blknos*[rw].dat).
1253 The -K option forces bno_plot.py to put the keys below the
1254 graph, typically all keys for input files are put in the
1255 upper right corner of the graph. If the number of devices
1256 exceed 10, then bno_plot.py will automatically push the
1257 keys under the graph.
1259 To exit the plotter, enter 'quit' or ^D at the 'gnuplot> '
1265 \leavevmode\centering
1266 \epsfig{file=bno_plot.eps,width=5.5in}
1267 \caption{\label{fig:bno_plot}Sample \texttt{bno\_plot.py} Screen Shot}
1271 \newpage\section{\label{sec:appendix}Sample \texttt{btt}
1273 Here is a complete output file from a btt run, illustrating a lot of the
1274 capabilities of btt.
1275 \input{sample-btt-output.tex}
1278 \subsection{\label{sec:o-B}\texttt{--dump-blocknos}/\texttt{-B}}