Flame Graph Reporter#
The flame graph reporter generates an HTML file containing a flame graph representation of the allocations contributing to the tracked process’s peak memory usage.
Flame graphs are a way to visualize how your program is spending its time. A few important things about flame graphs:
The flame graph displays the superposition of all stack traces that lead to memory allocations at a given time (normally the time when the total amount of allocated memory was highest).
A flame graph can’t tell you how many times a function was called, only how much memory that function allocated.
A stack trace is represented as a column of boxes, where each box represents a function call.
The x-axis does not show the passage of time, so the left-to-right ordering has no special meaning: every level just shows the collection of functions that were called by the nodes immediately under them.
The y-axis shows the stack depth, ordered from root at the bottom to leaf at the top. The top box shows the function that made a memory allocation and everything beneath that is its ancestry. The function beneath a function is its parent.
The width of each function box represents how much memory was allocated by that function or its children. Functions with wider boxes allocated more bytes of memory than those with narrower boxes, in proportion to their width.
Interpreting flame graphs#
Flame graphs can be interpreted as follows:
The nodes at the top of the flame graph represent functions that allocated memory.
For quickly identifying the functions that allocated more memory directly, look for large plateaus along the top edge, as these show a single stack trace was responsible for a large chunk of the total memory of the snapshot that the flame graph represents.
Reading the flame graph from top down shows the ancestry relationship: Every function is called by its parent, which is shown directly below it; the parent was called by its parent shown below it, and so on. A quick scan downward from a function identifies how it was called.
Reading the flame graph from bottom up shows code flow and the bigger picture. A function calls all child functions shown above it, which, in turn, call functions shown above them. Reading bottom up also shows the big picture of code flow before various forks split execution into smaller towers.
You can directly compare the width of function boxes as wider boxes mean more memory was allocated by the given node and those are the most important to understand first.
Major forks in the flame graph (when a node splits into several ones in the next level) can be useful to study: these nodes can indicate a logical grouping of code, where a function processes work in stages, each with its own function. It can also be caused by a conditional statement, which chooses which function to call.
If the application is multi-threaded, the stacks of all the threads that contribute to the memory peak will appear commingled in the flame graph.
def a(n): return b(n) def b(n): return [c(n), d(n)] def c(n): return "a" * n def d(n): return "a" * n a(100000)
This code allocates memory from the system allocator in just 2 places:
d(). This is how the flame graph looks:
Here you can see that
b() and that
d(), which in turn did some allocations. As the boxes of
d() are of the same width, you know that both allocated
the same amount of memory.
A more complete example#
def a(n): return [b(n), h(n)] def b(n): return c(n) def c(n): missing(n) return d(n) def missing(n): return "a" * n def d(n): return [e(n), f(n), "a" * (n // 2)] def e(n): return "a" * n def f(n): return g(n) def g(n): return "a" * n * 2 def h(n): return i(n) def i(n): return "a" * n a(100000)
This code allocates memory from the system allocator in 5 places:
missing(). The associated
flame graph looks like this:
The top edge shows that function
g() allocates the most memory,
d() is wider, but its exposed top edge is smaller, which means that
d() itself allocated less memory than the one allocated by the
functions called by it. Functions including
not allocate memory themselves directly; rather, the functions they
called did the allocating.
g() show its ancestry:
g() was called by
f(), which was called by
d(), and so on.
Visually comparing the widths of functions
b() code path allocated about four times more than
The actual functions that did the allocations in each case were their
A major fork in the code paths is visible where
h(). Understanding why the code does this may be a major clue to
its logical organization. This may be the result of a conditional (if
b(), else call
h()) or a logical grouping of
a() is processed in two parts:
In our case we know is the second case, as
a() is creating a list
with the result of
If you look carefully you can notice that
memory, but it does not appear in the flame graph. This is because at
the time the largest memory peak was reached (when
a() returned) the
memory allocated by
missing() didn’t contribute at all to the total
amount of memory. This is because the memory allocated by
is deallocated as soon as the call ends.
With this information, we know that if you need to choose a place to start
looking for optimizations, you should start looking at
e() and then
i() (in that order) as these are the places that
allocated the most memory when the program reached its maximum. Of
course, the actual optimization may be done in the callers of these
functions, but you have a way to start understanding where to optimize.
Non-relevant frame hiding#
The flame graph exposes a button to show or hide frames which might be distracting when interpreting the results, either because they were injected by Memray or because they are low-level implementation details of CPython. By default, frames tagged as non-relevant are hidden. You can reveal them with the “Show Non-Relevant Frames” checkbox:
Note that allocations in these frames will still be accounted for in parent frames, even if they’re hidden.
Memory Leaks View#
When generating flame graphs, the
--leaks option can be specified
to get information for memory that was leaked during the tracking
(i.e. allocated after tracking started and not deallocated by the time
The Python allocator doesn’t necessarily release memory to the system
when Python objects are deallocated and these can still appear as
“leaks”. When you use the
--leaks option, you should usually
also run your application with the
environment variable set. See our documentation on python
allocators for details.
When generating flame graphs, the
--split-threads option can be
specified to get thread-specific filtering on the flame graph.
--split-threads is not specified, thread information is not
displayed on the flame graph. Instead, allocations occurring at the same
source location across different threads are grouped together. However,
--split-threads is used, the allocation patterns of individual
threads can be analyzed.
When opening the report, the same merged thread view is presented, but a new “Filter Thread” dropdown will be shown. This can be used to select a specific thread to display a flame graph for that one thread:
To go back to the merged view, the “Reset” entry can be used in the dropdown menu.
Note that the root node (displayed as memray) is always present and is displayed as thread 0.
Flame graphs are effective visualization tools for a memory snapshot of a program. They give an insightful visual map for the execution of Python code and allow navigating areas of interest, letting you identify where to start looking for improvements. Unlike other code-path visualizations such as acyclic graphs, flame graphs convey information intuitively using line lengths and can handle large-scale profiles, while usually remaining readable on one screen.
The general form of the
flamegraph subcommand is:
memray flamegraph [options] <results>
The only argument the
flamegraph subcommand requires is the capture file
previously generated using the run subcommand.
The output file will be named
memray-flamegraph-<input file name>.html
-o argument was used to override the default name.
usage: memray flamegraph [-h] [-o OUTPUT] [-f] [--leaks | --temporary-allocation-threshold N | --temporary-allocations] [--split-threads] results
Results of the tracker run
- -o, --output
Output file name
- -f, --force
If the output file already exists, overwrite it
Enables Memory Leaks View, where memory that was not deallocated is displayed, instead of peak memory usage.
Report temporary allocations, as opposed to leaked allocations or high watermark allocations. An allocation is considered temporary if at most N other allocations occur before it is deallocated. With N=0, an allocation is temporary only if it is immediately deallocated before any other allocation occurs.
Equivalent to --temporary-allocation-threshold=1
Enables Split-Threads View, where each thread can be displayed separately. Allocations on the same source line across different threads are not merged, if this flag is passed.
Please submit feedback, ideas, and bug reports by filing a new issue at https://github.com/bloomberg/memray/issues