Skip to content

Commit

Permalink
Timeline visualization tool (#1022)
Browse files Browse the repository at this point in the history 
This PR adds eBPF-based scripts for recording the start and end time of
work packets and formatting the log for visualization with PerfettoUI.

Co-authored-by: Claire Huang <claire.x.huang@gmail.com>
Co-authored-by: Zixian Cai <u5937495@anu.edu.au>
  • Loading branch information
@wks @clairexhuang @caizixian
3 people authored Jan 9, 2024
1 parent e7143b9 commit b66fa35
Show file tree
Hide file tree
Showing 17 changed files with 675 additions and 207 deletions.
4 changes: 4 additions & 0 deletions Cargo.toml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -152,6 +152,10 @@ malloc_counted_size = []
# Count the size of all live objects in GC
count_live_bytes_in_gc = []

# Workaround a problem where bpftrace scripts (see tools/tracing/timeline/capture.bt) cannot
# capture the type names of work packets.
bpftrace_workaround = []

# Do not modify the following line - ci-common.sh matches it
# -- Mutally exclusive features --
# Only one feature from each group can be provided. Otherwise build will fail.
Expand Down
7 changes: 7 additions & 0 deletions src/scheduler/worker.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -384,6 +384,13 @@ impl<VM: VMBinding> GCWorker<VM> {
// probe! expands to an empty block on unsupported platforms
#[allow(unused_variables)]
let typename = work.get_type_name();

#[cfg(feature = "bpftrace_workaround")]
// Workaround a problem where bpftrace script cannot see the work packet names,
// by force loading from the packet name.
// See the "Known issues" section in `tools/tracing/timeline/README.md`
std::hint::black_box(unsafe { *(typename.as_ptr()) });

probe!(mmtk, work, typename.as_ptr(), typename.len());
work.do_work_with_stat(self, mmtk);
}
Expand Down
236 changes: 30 additions & 206 deletions tools/tracing/README.md
View file
Original file line number Diff line number Diff line change
@@ -1,230 +1,54 @@
# MMTk performace tracing
# eBPF-based tracing tools

## Notes for MMTk developers

Please open pull requests if you develop new tools that others might find useful.
When you add new tools, please update this documentation.
If you change MMTk internals that the tracing tools depend on (such as the
definition of `enum WorkBucketStage`), please update the scripts accordingly.

## Notes for MMTk users

Since some of the tools depend on the MMTk internals, please use the tools
shipped with the MMTk release you use.

## Tracepoints
Currently, the core provides the following tracepoints.
- `mmtk:collection_initialized()`: GC is enabled
- `mmtk:harness_begin()`: the timing iteration of a benchmark begins
- `mmtk:harness_end()`: the timing iteration of a benchmark ends
- `mmtk:gccontroller_run()`: the GC controller thread enters its work loop
- `mmtk:gcworker_run()`: a GC worker thread enters its work loop
- `mmtk:gc_start()`: a collection epoch starts
- `mmtk:gc_end()`: a collection epoch ends
- `mmtk:process_edges(num_edges: int, is_roots: bool)`: a invocation of the
`process_edges` method. The first argument is the number of edges to be processed,
and the second argument is whether these edges are root edges.
- `mmtk:bucket_opened(id: int)`: a work bucket opened. The first argument is the
numerical representation of `enum WorkBucketStage`.
- `mmtk:work_poll()`: a work packet is to be polled.
- `mmtk:work(type_name: char *, type_name_len: int)`: a work packet was just
executed. The first argument is points to the string of the Rust type name of
the work packet, and the second argument is the length of the string.
- `mmtk:alloc_slow_once_start()`: the allocation slow path starts.
- `mmtk:alloc_slow_once_end()`: the allocation slow path ends.

## Running tracing tools
The tracing tools are to be invoked by a wrapper script `run.py`.
```
usage: run.py [-h] [-b BPFTRACE] -m MMTK [-H] [-p] [-f {text,json}] tool

positional arguments:
tool Name of the bpftrace tool
optional arguments:
-h, --help show this help message and exit
-b BPFTRACE, --bpftrace BPFTRACE
Path of the bpftrace executable
-m MMTK, --mmtk MMTK Path of the MMTk binary
-H, --harness Only collect data for the timing iteration (harness_begin/harness_end)
-p, --print-script Print the content of the bpftrace script
-f {text,json}, --format {text,json}
bpftrace output format
```

- `-b`: the path to the `bpftrace` executable. By default, it uses `bpftrace`
executable in your `PATH`. We strongly recommend you use the latest statically
complied `bpftrace` from [upstream](https://github.com/iovisor/bpftrace/releases).
You need to be able to have sudo permission for whichever `bpftrace` you want to use.
- `-m`: the path to a MMTk binary that contains the tracepoints.
This depends on the binding you use.
For the OpenJDK binding, it should be `jdk/lib/server/libmmtk_openjdk.so` under
your build folder.
To check whether the binary contains tracepoints, you can use `readelf -n`.
You should see a bunch of `stapsdt` notes with `mmtk` as the provider.
- `-H`: pass this flag is you want to only measure the timing iteration of a
benchmark.
By default, the tracing tools will measure the entire execution.
- `-p`: print the entire tracing script before execution.
This is mainly for debugging use.
- `-f`: change the bpftrace output format.
By default, it uses human-readable plain text output (`text`).
You can set this to `json` for easy parsing.
Currently, the core provides the following tracepoints.

Please run the tracing tools **before** running the workload.
If you use `-H`, the tracing tools will automatically end with `harness_end` is
called.
Otherwise, you will need to terminate the tools manually with `Ctrl-C`.
These tools also have a timeout of 1200 seconds so not to stall unattended
benchmark execution.
- `mmtk:collection_initialized()`: GC is enabled
- `mmtk:harness_begin()`: the timing iteration of a benchmark begins
- `mmtk:harness_end()`: the timing iteration of a benchmark ends
- `mmtk:gccontroller_run()`: the GC controller thread enters its work loop
- `mmtk:gcworker_run()`: a GC worker thread enters its work loop
- `mmtk:gc_start()`: a collection epoch starts
- `mmtk:gc_end()`: a collection epoch ends
- `mmtk:process_edges(num_edges: int, is_roots: bool)`: a invocation of the `process_edges`
method. The first argument is the number of edges to be processed, and the second argument is
whether these edges are root edges.
- `mmtk:bucket_opened(id: int)`: a work bucket opened. The first argument is the numerical
representation of `enum WorkBucketStage`.
- `mmtk:work_poll()`: a work packet is to be polled.
- `mmtk:work(type_name: char *, type_name_len: int)`: a work packet was just executed. The first
argument is points to the string of the Rust type name of the work packet, and the second
argument is the length of the string.
- `mmtk:alloc_slow_once_start()`: the allocation slow path starts.
- `mmtk:alloc_slow_once_end()`: the allocation slow path ends.

## Tracing tools
### Measuring the time spend in allocation slow path (`alloc_slow`)
This tool measures the distribution of the allocation slow path time.
The time unit is 400ns, so that we use the histogram bins with higher
fidelity better.

Sample output:
```
@alloc_slow_hist:
[4, 8) 304 |@ |
[8, 16) 12603 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[16, 32) 8040 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32, 64) 941 |@@@ |
[64, 128) 171 | |
[128, 256) 13 | |
[256, 512) 2 | |
[512, 1K) 0 | |
[1K, 2K) 0 | |
[2K, 4K) 0 | |
[4K, 8K) 0 | |
[8K, 16K) 0 | |
[16K, 32K) 14 | |
[32K, 64K) 37 | |
[64K, 128K) 19 | |
[128K, 256K) 1 | |
```

In the above output, we can see that most allocation slow paths finish between
3.2us and 6.4us.
However, there is a long tail, presumably due to GC pauses.

### Measuring the time spend in different GC stages (`gc_stages`)
This tool measures the time spent in different stages of GC: before `Closure`,
during `Closure`, and after `Closure`.
The time unit is ns.

Sample output:
```
@closure_time: 1405302743
@post_closure_time: 81432919
@pre_closure_time: 103886118
```

In the above output, overall, the execution spends 1.4s in the main transitive
closure, 103ms before that, and 81ms after that (a total of around 1.5s).

### Measuring the time spend in lock contended state for Rust `Mutex` (`lock_contend`)
This tools measures the time spent in the lock contended state for Rust `Mutex`s.
The Rust standard library implements `Mutex` using the fast-slow-path paradigm.
Most lock operations take place in inlined fast paths, when there's no contention.
However, when there's contention,
`std::sys::unix::locks::futex_mutex::Mutex::lock_contended` is called.

```rust
#[inline]
pub fn lock(&self) {
if self.futex.compare_exchange(0, 1, Acquire, Relaxed).is_err() {
self.lock_contended();
}
}

#[cold]
fn lock_contended(&self) {
<snip>
}
```


MMTk uses Rust `Mutex`, e.g., in allocation slow paths for synchronization,
and this tool can be useful to measure the contention in these parts of code.

The time unit is 256ns.

Sample output:
```
@lock_dist[140637228007056]:
[1] 447 |@@@@ |
[2, 4) 3836 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[4, 8) 3505 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8, 16) 1354 |@@@@@@@@@@@@@@ |
[16, 32) 832 |@@@@@@@@ |
[32, 64) 1077 |@@@@@@@@@@@ |
[64, 128) 2991 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 4846 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[256, 512) 5013 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[512, 1K) 1203 |@@@@@@@@@@@@ |
[1K, 2K) 34 | |
[2K, 4K) 15 | |
```
Each sub-directory contains a set of scripts.

In the above output, we can see that the lock instance (140637228007056, or 0x7fe8a8047e90)
roughly has a bimodal distribution in terms of the time spent in lock contended
code path.
The first peak is around 512ns\~1024ns, and the second peak is around 66us\~131us.

If you can't tell which lock instance is for which lock in MMTk, you can trace
the allocation of the Mutex and record the stack trace (note that you might want
to compile MMTk with `force-frame-pointers` to obtain better stack traces).

### Measuring the distribution of `process_edges` packet sizes (`packet_size`)
Most of the GC time is spend in the transitive closure for tracing-based GCs,
and MMTk performs transitive closure via work packets that calls the `process_edges` method.
This tool measures the distribution of the sizes of these work packets, and also
count root edges separately.

Sample output:
```
@process_edges_packet_size:
[1] 238 |@@@@@ |
[2, 4) 806 |@@@@@@@@@@@@@@@@@ |
[4, 8) 1453 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8, 16) 1105 |@@@@@@@@@@@@@@@@@@@@@@@ |
[16, 32) 2410 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[32, 64) 1317 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[64, 128) 1252 |@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 1131 |@@@@@@@@@@@@@@@@@@@@@@@@ |
[256, 512) 2017 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[512, 1K) 1270 |@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[1K, 2K) 1028 |@@@@@@@@@@@@@@@@@@@@@@ |
[2K, 4K) 874 |@@@@@@@@@@@@@@@@@@ |
[4K, 8K) 1024 |@@@@@@@@@@@@@@@@@@@@@@ |
[8K, 16K) 58 |@ |
[16K, 32K) 5 | |
@process_edges_root_packet_size:
[1] 71 |@@@@@@@ |
[2, 4) 4 | |
[4, 8) 276 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8, 16) 495 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[16, 32) 477 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32, 64) 344 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[64, 128) 242 |@@@@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 109 |@@@@@@@@@@@ |
[256, 512) 31 |@@@ |
[512, 1K) 33 |@@@ |
[1K, 2K) 75 |@@@@@@@ |
[2K, 4K) 75 |@@@@@@@ |
[4K, 8K) 336 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8K, 16K) 56 |@@@@@ |
[16K, 32K) 3 | |
```

In the above output, we can see that overall, the sizes of the `process_edges`
has a unimodal distribution with a peak around 16\~32 edges per packet.
However, if we focus on root edges, the distribution is roughly bimodal, with a
first peak around 8\~16 and a second peak around 4096\~8192.
- `performance`: Print various GC-related statistics, such as the distribution of time spent in
allocation slow path, the time spent in each GC stages, and the distribution of `process_edges`
packet sizes.
- `timeline`: Record the start and end time of each GC and each work packet, and visualize them on
a timeline in Perfetto UI.

## Attribution

If used for research, please cite the following publication.

```bibtex
@inproceedings{DBLP:conf/pppj/HuangBC23,
author = {Claire Huang and
Expand Down
Loading

0 comments on commit b66fa35

Please sign in to comment.