Profile: Statistical Profiler
The profile collection implements a statistical profiler. The profiling is done by running a background thread that collects stack snapshots via continuation-mark-set->context, meaning that the result is an estimate of the execution costs and it is limited to the kind of information that continuation-mark-set->context produces (most notably being limited to functions calls, and subject to compiler optimizations); but the result is often useful. In practice, since this method does not require recompilation of your source and has very little runtime overhead, it can be used for longer runs which compensates for these limits.
1 Toplevel Interface
(require profile) |
This module provides one procedure and one macro that are convenient high-level entry points for profiling expressions. It abstracts over details that are available through other parts of the library, and is intended as a convenient tool for profiling code.
(profile-thunk thunk [ #:delay delay #:repeat iterations #:threads threads? #:render renderer #:periodic-renderer periodic-renderer]) → void? thunk : (-> any/c) delay : nonnegative-number? = 0.05 iterations : exact-nonnegative-integer? = 1 threads? : any/c = #f renderer : (profile? . -> . any/c) = text:render
periodic-renderer : (or/c #f (list/c nonnegative-number? (profile? . -> . any/c))) = #f
The profiler works by starting a “sampler” thread to periodically collect stack snapshots (using continuation-mark-set->context). To determine the frequency of these collections, the sampler thread sleeps delay seconds between collections. Note that this is will be close, but not identical to, the frequency in which data is actually sampled. (The delay value is passed on to create-sampler, which creates the sampler thread.)
Due to the statistical nature of the profiler, longer executions result in more accurate analysis. You can specify a number of iterations to repeat the thunk to collect more data.
Normally, the sampler collects only snapshots of the current-thread’s stack. Profiling a computation that creates threads will therefore lead to bad analysis: the timing results will be correct, but because the profiler is unaware of other threads the observed time will be suspiciously small, and work done in other threads will not be included in the results. To track all threads, specify a non-#f value for the threads? argument—
this will execute the computation in a fresh custodian, and keep track of all threads under this custodian. Once the computation is done and the sampler is stopped, the accumulated data is analyzed (by analyze-samples) and the resulting profile value is sent to the renderer function. See Profile Renderers for available renderers. You can also use values as a “renderer”—
in this case the profile-thunk returns the analyzed information which can now be rendered multiple times, or saved for future rendering. To provide feedback information during execution, specify a periodic-renderer. This should be a list holding a delay time (in seconds) and a renderer function. The delay determines the frequency in which the renderer is called, and it should be larger than the sampler delay (usually much larger since it can involve more noticeable overhead, and it is intended for a human observer).
(profile expr keyword-arguments ...)
2 Collecting Profile Information
(require profile/sampler) |
(create-sampler to-track delay [super-cust]) → ((symbol?) (any/c) . ->* . any/c)
to-track :
(or/c thread? custodian? (listof (or/c thread? custodian?))) delay : nonnegative-number? super-cust : custodian? = (current-custodian)
'pause and 'resume will stop or resume snapshot collection. These messages can be nested. Note that the thread will continue running—
it will just stop collecting snapshots. 'stop kills the sampler thread. It should be called when no additional data should be collected. (This is currently irreversible: there is no message to start a new sampler thread.)
'set-tracked! with a value will change the tracked object(s) which were initially specified as the to-track argument.
'set-tracked! with a numeric value will change the delay that the sampler is taking between snapshots. Note that although changing this means that the snapshots are not uniformly distributed, the results will still be correct: the cpu time between samples is taken into account when the collected data is analyzed.
Finally, a 'get-snapshots message will make the controller return the currently collected data. Note that this can be called multiple times, each call will return the data that is collected up to that point in time. In addition, it can be (and usually is) called after the sampler was stopped.
The value that is returned should be considered as an undocumented internal detail of the profiler, intended to be sent to analyze-samples for analysis. The reason this is not done automatically, is that a future extension might allow you to combine several sampler results, making it possible to combine a profile analysis from several individual runs, possibly from different machines.
3 Analyzing Profile Data
(require profile/analyzer) |
Once a profile run is done and the results are collected, the next step is to analyze the data. In this step sample times are computed and summed, a call-graph representing observed function calls is built, and per-node and per-edge information is created. This is the job of the main function provided by profile/analyzer.
(analyze-samples raw-sample-data) → profile? raw-sample-data : any/c
(struct profile ( total-time cpu-time sample-number thread-times nodes *-node)) total-time : exact-nonnegative-integer? cpu-time : exact-nonnegative-integer? sample-number : exact-nonnegative-integer?
thread-times :
(listof (cons exact-nonnegative-integer? exact-nonnegative-integer?)) nodes : (listof node?) *-node : node?
total-time is the total observed time (in milliseconds) included in the profile run. This can be different from the actual time the profiling took, due to unaccounted-for time spent in untracked threads. (E.g., time spent in the sampler thread itself.)
cpu-time is the actual cpu time consumed by the process during the profiler’s work.
sample-number holds the number of samples taken during the profile. This can be used to compute the average time frame each of the input samples represented.
thread-times holds an association list mapping thread identifiers to cpu time for the corresponding threads. As samples are collected, each thread that is observed is assigned a small integer identifier. These identifiers are listed for each function call, and the total time spent in each thread is in this field.
nodes is a list of nodes representing all observed functions. These nodes are the components of the call-graph that the analyzer assembles (see the edge field). The nodes are sorted by a topological top-to-bottom sort, and by decreasing total amount of time (time spent either in the function or in its callees) as a secondary key.
*-node holds a “special” root node value that is constructed for every call graph. This node is used as the caller for all top-level function nodes and as the callee for all leaf nodes. It can therefore be used to start a recursive scan of the call graph. In addition, the times associated with its “callers” and “callees” actually represent the time these functions spent being the root of the computation or its leaf. (This can be different from a node’s “self” time, since it is divided by the number of instances a function had on the stack in each sample—
so for recursive functions this value is always different from the “self” time.)
(struct node (id src thread-ids total self callers callees)) id : (or/c #f symbol?) src : (or/c #f srcloc?) thread-ids : (listof exact-nonnegative-integer?) total : exact-nonnegative-integer? self : exact-nonnegative-integer? callers : (listof edge?) callees : (listof edge?)
The id and src fields hold a symbol naming the function and/or its source location as a srcloc value. This is the same as the results of continuation-mark-set->context, so at most one of these can be #f, except for the special *-node (see the profile struct) that can be identified by both being #f.
thread-ids holds a list of thread identifiers that were observed executing this function.
total holds the total time (in milliseconds) that this function was anywhere on the stack. It is common to see a few toplevel functions that have close to a 100% total time, but otherwise small self times—
these functions are the ones that initiate the actual work, but they don’t do any hard work directly. self holds the total time (in milliseconds) that this function was observed as the leaf of the stack. It represents the actual work done by this function, rather than the total time spent by both the function and its callees.
callers and callees hold the list of callers and callees. The nodes are not actually held in these lists, instead, edge values are used—
and provide information specific to each edge in the call-graph.
(struct edge (total caller caller-time callee callee-time)) total : exact-nonnegative-integer? caller : node? caller-time : exact-nonnegative-integer? callee : node? callee-time : exact-nonnegative-integer?
total is analogous to the total field of a node value: the total time that this edge was anywhere on the stack.
caller and callee hold the two nodes that are connected by this edge.
caller-time and callee-time hold the time spent on this edge from the caller’s or the callee’s perspective. These times are different from each other (as well as from the total time) because the sums that make them are each divided by the number of times the caller or the callee was on the stack.
To understand this difference, consider a stack snapshot holding A → B → B → B → A, and representing a second of observed cpu time. For this sample, the A → B edge is charged by a whole second for its total time (the same goes for the A → A edge, for example). Its caller time is charged 1/2 second because A appears twice in this stack snapshot (in the other half, A is chared for being a leaf —
the caller of the special *-node), and its callee time is charged 1/3 respectively.
4 Profile Renderers
After collecting the profile samples and analyzing the data, the last step of the profiling process is to render the results. The profile collection provides several renderers, each providing a rendering function that consumes a profile instance. See the analyzer section for a description of the profile struct if you want to implement a new renderer.
4.1 Textual Rendering
(require profile/render-text) |
(render profile-data [ #:truncate-source truncate-source #:hide-self hide-self% #:hide-subs hide-subs%]) → void? profile-data : profile? truncate-source : exact-nonnegative-integer? = 50 hide-self% : (between/c 0 1) = 1/100 hide-subs% : (between/c 0 1) = 1/50
The printout begins with some general details about the profile, and then a table that represents the call-graph is printed. Each row in this table looks like:
B [M1] M2% |
[N1] N2(N3%) N4(N5%) A ...path/to/source.rkt:12:34 |
C [M3] M4% |
A —
the name of the function that this node represents, followed by the source location for the function if it is known. The name can be “???” for anonymous functions, but in this case the source location will identify them. N1 —
an index number associated with this node. This is useful in references to this function, since the symbolic names are not unique (and some can be missing). The number itself has no significance. N2 —
the time (in milliseconds) that this function has been anywhere in a stack snapshot. This is the total time that the execution was somewhere in this function or in its callees. (Corresponds to the node-total field.) N3 —
this is the percentage of the node’s total time (N2) from the total observed time of the profile. An entry with a 100% refers to a function that was active throughout the whole execution. N4 —
the time (in milliseconds) that this function has been at the top of the stack snapshot. This is the time that this function was itself doing work rather than calling other functions. (Corresponds to the node-self field.) N5 —
this is the percentage of N4 out of the total observed time of the profile. Functions with high values here can be good candidates for optimization, But, of course, they can represent doing real work for a caller that needs to be optimized. B and C —
these are labels for the callers and callees of the function. Any number of callers and callees can appear here (including 0). The function itself can also appear in both places if it is (non-tail) recursive. M1 and M3 —
the index numbers for B and C. They can be used to disambiguate functions with the same name, as well as a quick way to find the corresponding entry in the table. M2 and M4 —
the percentages of the time A spent being called by B and calling C. These percentages represent the time that this edge was found on a stack snapshot, divided by the number of occurrences of A on the same snapshot. The number is the percentage of these times out of N2, the total time A has been active. The total percentages for the all caller and for all callees should be close to 100% minus the time A was the leaf or the root.
These values correspond to the edge-caller-time and edge-callee-time fields; see the documentation for further details.
The truncate-source argument determines the length that the source string should take (together with its label).
hide-self% and hide-subs% control hiding some of the nodes. A node is hidden if its self time (N3 in the above example) is smaller than hide-self% and if all places where it occurs as a caller or a callee have percentages that are smaller than hide-subs%. The reason for requiring both conditions is to avoid having “dangling references” to hidden nodes.
4.2 Graph Rendering
(require profile/render-graphviz) |
(render profile-data [ #:hide-self hide-self% #:hide-subs hide-subs%]) → void? profile-data : profile? hide-self% : (between/c 0 1) = 1/100 hide-subs% : (between/c 0 1) = 1/50
This is an experimental module, provided mostly as a proof-of-concept. It renders the profile’s call-graph as a graph representation for one of the Graphviz tools to render. Nodes are colored according to their `self’ percentages, and edges.
The keyword arguments control hiding nodes in the same way as with the textual renderer.