midden

Find what is holding a JVM heap dump's memory: retained sizes, owners, leak suspects. One static binary, no JDK needed.

It reads .hprof dumps from jcmd, jmap or -XX:+HeapDumpOnOutOfMemoryError, gzipped or inside a tar or zip, and builds the full object graph and dominator tree, so every object has a retained size.

Install

Every version is on the releases page with its changes and binaries for Linux (static, any distro), Windows and macOS ARM64 (M-series).

curl -fsSLO https://github.com/NullByte3/midden/releases/latest/download/midden-x86_64-unknown-linux-musl
sudo install midden-x86_64-unknown-linux-musl /usr/local/bin/midden

curl -fsSLO https://github.com/NullByte3/midden/releases/latest/download/midden-aarch64-apple-darwin
sudo install -d /usr/local/bin
sudo install midden-aarch64-apple-darwin /usr/local/bin/midden

Invoke-WebRequest https://github.com/NullByte3/midden/releases/latest/download/midden-x86_64-pc-windows-gnu.exe -OutFile midden.exe

Or cargo install --path ..

Usage

midden server.hprof                          # full report
midden server.hprof --class ObjectMapper     # one class: referrers, owners, root paths
midden server.hprof --object 0x7f3a1234      # one object: fields, owners, what it retains
midden server.hprof --find jdbc:postgresql   # strings containing the text, and who holds them
midden server.hprof --where 'StandardSession.isValid=false'
midden after.hprof --baseline before.hprof   # what grew between two dumps
midden server.hprof --only suspects --json   # for scripts
midden server.hprof --shell                  # drill in without re-reading the file

  leak suspects  objects or classes retaining ≥10% of the live heap
    1.  983.7MB  64.1%  class Big @0x740920000
          979.5MB  63.8%  static Big.MAP java.util.HashMap @0x740920090
          979.5MB  63.8%  .table java.util.HashMap$Node[8,388,608] @0x77ba00000  ← accumulation point
       keeps  27,999,745 objects: 4,000,000 × int[] 183.1MB | 4,000,000 × HashMap$Node 183.1MB | ...

Sections: heap, suspects, biggest, classes, collections, threads, locals, loaders, strings, arrays, boxed, references, garbage, system, direct and baseline. Pick them with --only and --skip. midden --help lists every flag.

• Soft, weak and phantom referents do not count as held unless --include-soft or --include-weak.
• Sizes follow MAT by default. --sizes compressed gives HotSpot's compressed-oops footprint.
• The first run caches the index in <dump>.midden/, so later runs start at the report. Turn it off with --no-cache.
• Expect about 90 bytes of RAM per object. The dump itself is streamed.

Benchmark

midden 0.2.0 compared with Eclipse MAT 1.17, JProfiler 16.2.1 and the VisualVM 2.2.2 heap engine, run on 8 CPUs with 4 GiB of memory. Each dump was in the page cache before its runs, and the tables show the best of 2 or 3 runs. The other tools were measured alongside midden 0.1.0 on the same machine.

Cold first run

Time from start to a finished result, with no index or cache on disk. MAT is parse plus Leak Suspects. JProfiler is jpanalyze -retained=true, which builds the index its GUI opens but writes no report.

Dump	Objects	midden	midden, 1 thread	MAT	JProfiler	VisualVM
Paper server, 938 MB	11.2 M	1.30 seconds	3.53 seconds	20.95 seconds	12.05 seconds	55.8 seconds
mixed-small, 414 MB	4.1 M	0.40 seconds	1.11 seconds	7.09 seconds	4.11 seconds	772 seconds
mixed-1g, 1.2 GB	11.8 M	1.10 seconds	3.78 seconds	17.15 seconds	10.25 seconds	over 1,800 seconds, stopped
arrays-1g, 1.2 GB	37 k	0.08 seconds	0.26 seconds	1.84 seconds	0.97 seconds	0.37 seconds
chain-8m, 429 MB	8.0 M	0.85 seconds	1.64 seconds	12.26 seconds	7.01 seconds	23.8 seconds
small-20m, 945 MB	19.2 M	5.85 seconds	14.71 seconds	45.71 seconds	33.46 seconds	334 seconds

midden was 8 to 23 times faster than MAT and 6 to 12 times faster than JProfiler, and it was faster on a single thread than either of them on eight. MAT's full three-report run (suspects, overview and top components) took 26 to 38 seconds on Paper.

Reopening an indexed dump

Dump	midden	MAT	JProfiler
Paper	1.06 seconds	30.59 seconds	1.54 seconds
small-20m	5.49 seconds	18.80 seconds	3.94 seconds

JProfiler reopens small-20m faster because its analysis stores the dominator tree. midden's cache only stores the parse, so it saves 0.2 to 0.4 seconds.

Memory

Peak RSS on Paper: midden 0.6 GB, JProfiler 1.3 GB, MAT 2.1 to 3.3 GB (at its 3 GB heap), VisualVM 3.5 GB.

Features

	midden	MAT	JProfiler	VisualVM
Retained sizes and dominator tree	yes	yes	yes	yes
Leak suspects report	yes, objects and classes, with the accumulation point	yes	no, biggest objects view	no
Path from a GC root, with field names	yes, printed for every suspect	yes	yes, in the GUI	yes, in the GUI
Thread, frame and line that roots an object	yes	yes	in the GUI	in the GUI
Class histogram, shallow and retained	yes, also by package	yes	yes	yes
Collection fill ratios and empty collections	yes	yes	yes, inspections	through OQL
Duplicate strings and arrays	yes	yes	yes, inspections	through OQL
Thread locals, class loaders	yes	yes	class loaders	not checked
Soft, weak and phantom references	excluded by default, --include-soft and --include-weak	followed	excluded, options to keep	followed
Unreachable objects	reported separately	dropped, option to keep	dropped, option to keep	not checked
Direct buffers	yes	through OQL	not checked	through OQL
Compare two dumps	--baseline	histogram compare in the GUI	jpcompare and the GUI	yes, in the GUI
Query objects	--class, --object, --find, --where, --shell	OQL	heap walker filters	OQL
Reads gzip, tar and zip	yes	gzip	no	not checked
Headless report	yes, text or --json	yes, HTML zips	no, index only, results over MCP	no
Index cache	yes	yes	yes	yes
Needs a JVM	no, one static binary	yes	yes	yes
License	MIT	EPL 2.0	commercial	GPL 2 with Classpath Exception

Development

CI runs cargo fmt --check and cargo clippy --all-targets -- -D warnings.

License

MIT, see LICENSE.