/mimalloc

mimalloc is a compact general purpose allocator with excellent performance.

Primary LanguageCMIT LicenseMIT

mimalloc

 

mimalloc (pronounced "me-malloc") is a general purpose allocator with excellent performance characteristics. Initially developed by Daan Leijen for the run-time systems of the Koka and Lean languages. Latest release:v1.6.2 (2020-04-20).

It is a drop-in replacement for malloc and can be used in other programs without code changes, for example, on dynamically linked ELF-based systems (Linux, BSD, etc.) you can use it as:

> LD_PRELOAD=/usr/bin/libmimalloc.so  myprogram

It also has an easy way to override the allocator in Windows. Notable aspects of the design include:

  • small and consistent: the library is about 6k LOC using simple and consistent data structures. This makes it very suitable to integrate and adapt in other projects. For runtime systems it provides hooks for a monotonic heartbeat and deferred freeing (for bounded worst-case times with reference counting).
  • free list sharding: the big idea: instead of one big free list (per size class) we have many smaller lists per memory "page" which both reduces fragmentation and increases locality -- things that are allocated close in time get allocated close in memory. (A memory "page" in mimalloc contains blocks of one size class and is usually 64KiB on a 64-bit system).
  • eager page reset: when a "page" becomes empty (with increased chance due to free list sharding) the memory is marked to the OS as unused ("reset" or "purged") reducing (real) memory pressure and fragmentation, especially in long running programs.
  • secure: mimalloc can be built in secure mode, adding guard pages, randomized allocation, encrypted free lists, etc. to protect against various heap vulnerabilities. The performance penalty is usually around 10% on average over our benchmarks.
  • first-class heaps: efficiently create and use multiple heaps to allocate across different regions. A heap can be destroyed at once instead of deallocating each object separately.
  • bounded: it does not suffer from blowup [1], has bounded worst-case allocation times (wcat), bounded space overhead (~0.2% meta-data, with at most 12.5% waste in allocation sizes), and has no internal points of contention using only atomic operations.
  • fast: In our benchmarks (see below), mimalloc outperforms other leading allocators (jemalloc, tcmalloc, Hoard, etc), and usually uses less memory (up to 25% more in the worst case). A nice property is that it does consistently well over a wide range of benchmarks. There is also good huge OS page support for larger server programs.

The documentation gives a full overview of the API. You can read more on the design of mimalloc in the technical report which also has detailed benchmark results.

Enjoy!

Releases

  • 2020-04-20, v1.6.2: stable release 1.6: fix compilation on Android, MingW, Raspberry, and Conda, stability fix for Windows 7, fix multiple mimalloc instances in one executable, fix strnlen overload, fix aligned debug padding.
  • 2020-02-17, v1.6.1: stable release 1.6: minor updates (build with clang-cl, fix alignment issue for small objects).
  • 2020-02-09, v1.6.0: stable release 1.6: fixed potential memory leak, improved overriding and thread local support on FreeBSD, NetBSD, DragonFly, and macOSX. New byte-precise heap block overflow detection in debug mode (besides the double-free detection and free-list corruption detection). Add nodiscard attribute to most allocation functions. Enable MIMALLOC_PAGE_RESET by default. New reclamation strategy for abandoned heap pages for better memory footprint.
  • 2020-02-09, v1.5.0: stable release 1.5: improved free performance, small bug fixes.
  • 2020-01-22, v1.4.0: stable release 1.4: improved performance for delayed OS page reset, more eager concurrent free, addition of STL allocator, fixed potential memory leak.
  • 2020-01-15, v1.3.0: stable release 1.3: bug fixes, improved randomness and stronger free list encoding in secure mode.
  • 2019-12-22, v1.2.2: stable release 1.2: minor updates.
  • 2019-11-22, v1.2.0: stable release 1.2: bug fixes, improved secure mode (free list corruption checks, double free mitigation). Improved dynamic overriding on Windows.
  • 2019-10-07, v1.1.0: stable release 1.1.
  • 2019-09-01, v1.0.8: pre-release 8: more robust windows dynamic overriding, initial huge page support.
  • 2019-08-10, v1.0.6: pre-release 6: various performance improvements.

Special thanks to:

  • Jason Gibson (@jasongibson) for exhaustive testing on large workloads and server environments and finding complex bugs in (early versions of) mimalloc.
  • Manuel Pöter (@mpoeter) and Sam Gross (@colesbury) for finding an ABA concurrency issue in abandoned segment reclamation.

Building

Windows

Open ide/vs2019/mimalloc.sln in Visual Studio 2019 and build (or ide/vs2017/mimalloc.sln). The mimalloc project builds a static library (in out/msvc-x64), while the mimalloc-override project builds a DLL for overriding malloc in the entire program.

macOS, Linux, BSD, etc.

We use cmake1 as the build system:

> mkdir -p out/release
> cd out/release
> cmake ../..
> make

This builds the library as a shared (dynamic) library (.so or .dylib), a static library (.a), and as a single object file (.o).

> sudo make install (install the library and header files in /usr/local/lib and /usr/local/include)

You can build the debug version which does many internal checks and maintains detailed statistics as:

> mkdir -p out/debug
> cd out/debug
> cmake -DCMAKE_BUILD_TYPE=Debug ../..
> make

This will name the shared library as libmimalloc-debug.so.

Finally, you can build a secure version that uses guard pages, encrypted free lists, etc., as:

> mkdir -p out/secure
> cd out/secure
> cmake -DMI_SECURE=ON ../..
> make

This will name the shared library as libmimalloc-secure.so. Use ccmake2 instead of cmake to see and customize all the available build options.

Notes:

  1. Install CMake: sudo apt-get install cmake
  2. Install CCMake: sudo apt-get install cmake-curses-gui

Using the library

The preferred usage is including <mimalloc.h>, linking with the shared- or static library, and using the mi_malloc API exclusively for allocation. For example,

> gcc -o myprogram -lmimalloc myfile.c

mimalloc uses only safe OS calls (mmap and VirtualAlloc) and can co-exist with other allocators linked to the same program. If you use cmake, you can simply use:

find_package(mimalloc 1.4 REQUIRED)

in your CMakeLists.txt to find a locally installed mimalloc. Then use either:

target_link_libraries(myapp PUBLIC mimalloc)

to link with the shared (dynamic) library, or:

target_link_libraries(myapp PUBLIC mimalloc-static)

to link with the static library. See test\CMakeLists.txt for an example.

For best performance in C++ programs, it is also recommended to override the global new and delete operators. For convience, mimalloc provides mimalloc-new-delete.h which does this for you -- just include it in a single(!) source file in your project. In C++, mimalloc also provides the mi_stl_allocator struct which implements the std::allocator interface.

You can pass environment variables to print verbose messages (MIMALLOC_VERBOSE=1) and statistics (MIMALLOC_SHOW_STATS=1) (in the debug version):

> env MIMALLOC_SHOW_STATS=1 ./cfrac 175451865205073170563711388363

175451865205073170563711388363 = 374456281610909315237213 * 468551

heap stats:     peak      total      freed       unit
normal   2:    16.4 kb    17.5 mb    17.5 mb      16 b   ok
normal   3:    16.3 kb    15.2 mb    15.2 mb      24 b   ok
normal   4:      64 b      4.6 kb     4.6 kb      32 b   ok
normal   5:      80 b    118.4 kb   118.4 kb      40 b   ok
normal   6:      48 b       48 b       48 b       48 b   ok
normal  17:     960 b      960 b      960 b      320 b   ok

heap stats:     peak      total      freed       unit
    normal:    33.9 kb    32.8 mb    32.8 mb       1 b   ok
      huge:       0 b        0 b        0 b        1 b   ok
     total:    33.9 kb    32.8 mb    32.8 mb       1 b   ok
malloc requested:         32.8 mb

 committed:    58.2 kb    58.2 kb    58.2 kb       1 b   ok
  reserved:     2.0 mb     2.0 mb     2.0 mb       1 b   ok
     reset:       0 b        0 b        0 b        1 b   ok
  segments:       1          1          1
-abandoned:       0
     pages:       6          6          6
-abandoned:       0
     mmaps:       3
 mmap fast:       0
 mmap slow:       1
   threads:       0
   elapsed:     2.022s
   process: user: 1.781s, system: 0.016s, faults: 756, reclaims: 0, rss: 2.7 mb

The above model of using the mi_ prefixed API is not always possible though in existing programs that already use the standard malloc interface, and another option is to override the standard malloc interface completely and redirect all calls to the mimalloc library instead .

Environment Options

You can set further options either programmatically (using mi_option_set), or via environment variables.

  • MIMALLOC_SHOW_STATS=1: show statistics when the program terminates.
  • MIMALLOC_VERBOSE=1: show verbose messages.
  • MIMALLOC_SHOW_ERRORS=1: show error and warning messages.
  • MIMALLOC_PAGE_RESET=0: by default, mimalloc will reset (or purge) OS pages that are not in use, to signal to the OS that the underlying physical memory can be reused. This can reduce memory fragmentation in long running (server) programs. By setting it to 0 this will no longer be done which can improve performance for batch-like programs. As an alternative, the MIMALLOC_RESET_DELAY= can be set higher (100ms by default) to make the page reset occur less frequently instead of turning it off completely.
  • MIMALLOC_USE_NUMA_NODES=N: pretend there are at most N NUMA nodes. If not set, the actual NUMA nodes are detected at runtime. Setting N to 1 may avoid problems in some virtual environments. Also, setting it to a lower number than the actual NUMA nodes is fine and will only cause threads to potentially allocate more memory across actual NUMA nodes (but this can happen in any case as NUMA local allocation is always a best effort but not guaranteed).
  • MIMALLOC_LARGE_OS_PAGES=1: use large OS pages (2MiB) when available; for some workloads this can significantly improve performance. Use MIMALLOC_VERBOSE to check if the large OS pages are enabled -- usually one needs to explicitly allow large OS pages (as on Windows and Linux). However, sometimes the OS is very slow to reserve contiguous physical memory for large OS pages so use with care on systems that can have fragmented memory (for that reason, we generally recommend to use MIMALLOC_RESERVE_HUGE_OS_PAGES instead whenever possible).
  • MIMALLOC_RESERVE_HUGE_OS_PAGES=N: where N is the number of 1GiB huge OS pages. This reserves the huge pages at startup and sometimes this can give a large (latency) performance improvement on big workloads. Usually it is better to not use MIMALLOC_LARGE_OS_PAGES in combination with this setting. Just like large OS pages, use with care as reserving contiguous physical memory can take a long time when memory is fragmented (but reserving the huge pages is done at startup only once). Note that we usually need to explicitly enable huge OS pages (as on Windows and Linux)). With huge OS pages, it may be beneficial to set the setting MIMALLOC_EAGER_COMMIT_DELAY=N (N is 1 by default) to delay the initial N segments (of 4MiB) of a thread to not allocate in the huge OS pages; this prevents threads that are short lived and allocate just a little to take up space in the huge OS page area (which cannot be reset).

Use caution when using fork in combination with either large or huge OS pages: on a fork, the OS uses copy-on-write for all pages in the original process including the huge OS pages. When any memory is now written in that area, the OS will copy the entire 1GiB huge page (or 2MiB large page) which can cause the memory usage to grow in big increments.

Overriding Malloc

Overriding the standard malloc can be done either dynamically or statically.

Dynamic override

This is the recommended way to override the standard malloc interface.

Override on Linux, BSD

On these ELF-based systems we preload the mimalloc shared library so all calls to the standard malloc interface are resolved to the mimalloc library.

> env LD_PRELOAD=/usr/lib/libmimalloc.so myprogram

You can set extra environment variables to check that mimalloc is running, like:

> env MIMALLOC_VERBOSE=1 LD_PRELOAD=/usr/lib/libmimalloc.so myprogram

or run with the debug version to get detailed statistics:

> env MIMALLOC_SHOW_STATS=1 LD_PRELOAD=/usr/lib/libmimalloc-debug.so myprogram

Override on MacOS

On macOS we can also preload the mimalloc shared library so all calls to the standard malloc interface are resolved to the mimalloc library.

> env DYLD_FORCE_FLAT_NAMESPACE=1 DYLD_INSERT_LIBRARIES=/usr/lib/libmimalloc.dylib myprogram

Note that certain security restrictions may apply when doing this from the shell.

(Note: macOS support for dynamic overriding is recent, please report any issues.)

Override on Windows

Overriding on Windows is robust and has the particular advantage to be able to redirect all malloc/free calls that go through the (dynamic) C runtime allocator, including those from other DLL's or libraries.

The overriding on Windows requires that you link your program explicitly with the mimalloc DLL and use the C-runtime library as a DLL (using the /MD or /MDd switch). Also, the mimalloc-redirect.dll (or mimalloc-redirect32.dll) must be available in the same folder as the main mimalloc-override.dll at runtime (as it is a dependency). The redirection DLL ensures that all calls to the C runtime malloc API get redirected to mimalloc (in mimalloc-override.dll).

To ensure the mimalloc DLL is loaded at run-time it is easiest to insert some call to the mimalloc API in the main function, like mi_version() (or use the /INCLUDE:mi_version switch on the linker). See the mimalloc-override-test project for an example on how to use this. For best performance on Windows with C++, it is also recommended to also override the new/delete operations (by including mimalloc-new-delete.h a single(!) source file in your project).

The environment variable MIMALLOC_DISABLE_REDIRECT=1 can be used to disable dynamic overriding at run-time. Use MIMALLOC_VERBOSE=1 to check if mimalloc was successfully redirected.

(Note: in principle, it is possible to even patch existing executables without any recompilation if they are linked with the dynamic C runtime (ucrtbase.dll) -- just put the mimalloc-override.dll into the import table (and put mimalloc-redirect.dll in the same folder) Such patching can be done for example with CFF Explorer).

Static override

On Unix-like systems, you can also statically link with mimalloc to override the standard malloc interface. The recommended way is to link the final program with the mimalloc single object file (mimalloc-override.o). We use an object file instead of a library file as linkers give preference to that over archives to resolve symbols. To ensure that the standard malloc interface resolves to the mimalloc library, link it as the first object file. For example:

> gcc -o myprogram mimalloc-override.o  myfile1.c ...

Another way to override statically that works on all platforms, is to link statically to mimalloc (as shown in the introduction) and include a header file in each source file that re-defines malloc etc. to mi_malloc. This is provided by mimalloc-override.h. This only works reliably though if all sources are under your control or otherwise mixing of pointers from different heaps may occur!

Performance

Last update: 2020-01-20

We tested mimalloc against many other top allocators over a wide range of benchmarks, ranging from various real world programs to synthetic benchmarks that see how the allocator behaves under more extreme circumstances. In our benchmark suite, mimalloc outperforms other leading allocators (jemalloc, tcmalloc, Hoard, etc), and has a similar memory footprint. A nice property is that it does consistently well over the wide range of benchmarks.

General memory allocators are interesting as there exists no algorithm that is optimal -- for a given allocator one can usually construct a workload where it does not do so well. The goal is thus to find an allocation strategy that performs well over a wide range of benchmarks without suffering from (too much) underperformance in less common situations.

As always, interpret these results with care since some benchmarks test synthetic or uncommon situations that may never apply to your workloads. For example, most allocators do not do well on xmalloc-testN but that includes the best industrial allocators like jemalloc and tcmalloc that are used in some of the world's largest systems (like Chrome or FreeBSD).

We show here only an overview -- for more specific details and further benchmarks we refer to the technical report. The benchmark suite is automated and available separately as mimalloc-bench.

Benchmark Results on 36-core Intel

Testing on a big Amazon EC2 compute instance (c5.18xlarge) consisting of a 72 processor Intel Xeon at 3GHz with 144GiB ECC memory, running Ubuntu 18.04.1 with LibC 2.27 and GCC 7.4.0. The measured allocators are mimalloc (xmi, tag:v1.4.0, page reset enabled) and its secure build as smi, Google's tcmalloc (tc, tag:gperftools-2.7) used in Chrome, Facebook's jemalloc (je, tag:5.2.1) by Jason Evans used in Firefox and FreeBSD, the Intel thread building blocks allocator (tbb, tag:2020), rpmalloc (rp,tag:1.4.0) by Mattias Jansson, the original scalable Hoard (tag:3.13) allocator by Emery Berger [1], the memory compacting Mesh (git:51222e7) allocator by Bobby Powers et al [8], and finally the default system allocator (glibc, 2.7.0) (based on PtMalloc2).

Any benchmarks ending in N run on all processors in parallel. Results are averaged over 10 runs and reported relative to mimalloc (where 1.2 means it took 1.2× longer to run). The legend also contains the overall relative score between the allocators where 100 points is the maximum if an allocator is fastest on all benchmarks.

The single threaded cfrac benchmark by Dave Barrett is an implementation of continued fraction factorization which uses many small short-lived allocations. All allocators do well on such common usage, where mimalloc is just a tad faster than tcmalloc and jemalloc.

The leanN program is interesting as a large realistic and concurrent workload of the Lean theorem prover compiling its own standard library, and there is a 7% speedup over tcmalloc. This is quite significant: if Lean spends 20% of its time in the allocator that means that mimalloc is 1.3× faster than tcmalloc here. (This is surprising as that is not measured in a pure allocation benchmark like alloc-test. We conjecture that we see this outsized improvement here because mimalloc has better locality in the allocation which improves performance for the other computations in a program as well).

The single threaded redis benchmark again show that most allocators do well on such workloads where tcmalloc did best this time.

The larsonN server benchmark by Larson and Krishnan [2] allocates and frees between threads. They observed this behavior (which they call bleeding) in actual server applications, and the benchmark simulates this. Here, mimalloc is quite a bit faster than tcmalloc and jemalloc probably due to the object migration between different threads.

The mstressN workload performs many allocations and re-allocations, and migrates objects between threads (as in larsonN). However, it also creates and destroys the N worker threads a few times keeping some objects alive beyond the life time of the allocating thread. We observed this behavior in many larger server applications.

The rptestN benchmark by Mattias Jansson is a allocator test originally designed for rpmalloc, and tries to simulate realistic allocation patterns over multiple threads. Here the differences between allocators become more apparent.

The second benchmark set tests specific aspects of the allocators and shows even more extreme differences between them.

The alloc-test, by OLogN Technologies AG, is a very allocation intensive benchmark doing millions of allocations in various size classes. The test is scaled such that when an allocator performs almost identically on alloc-test1 as alloc-testN it means that it scales linearly. Here, tcmalloc, and Hoard seem to scale less well and do more than 10% worse on the multi-core version. Even the best industrial allocators (tcmalloc, jemalloc, and tbb) are more than 10% slower as mimalloc here.

The sh6bench and sh8bench benchmarks are developed by MicroQuill as part of SmartHeap. In sh6bench mimalloc does much better than the others (more than 1.5× faster than jemalloc). We cannot explain this well but believe it is caused in part by the "reverse" free-ing pattern in sh6bench. The sh8bench is a variation with object migration between threads; whereas tcmalloc did well on sh6bench, the addition of object migration causes it to be 10× slower than before.

The xmalloc-testN benchmark by Lever and Boreham [5] and Christian Eder, simulates an asymmetric workload where some threads only allocate, and others only free -- they observed this pattern in larger server applications. Here we see that the mimalloc technique of having non-contended sharded thread free lists pays off as it outperforms others by a very large margin. Only rpmalloc and tbb also scale well on this benchmark.

The cache-scratch benchmark by Emery Berger [1], and introduced with the Hoard allocator to test for passive-false sharing of cache lines. With a single thread they all perform the same, but when running with multiple threads the potential allocator induced false sharing of the cache lines can cause large run-time differences. Crundal [6] describes in detail why the false cache line sharing occurs in the tcmalloc design, and also discusses how this can be avoided with some small implementation changes. Only the tbb, rpmalloc and mesh allocators also avoid the cache line sharing completely, while Hoard and glibc seem to mitigate the effects. Kukanov and Voss [7] describe in detail how the design of tbb avoids the false cache line sharing.

On 24-core AMD Epyc

For completeness, here are the results on a r5a.12xlarge instance having a 48 processor AMD Epyc 7000 at 2.5GHz with 384GiB of memory. The results are similar to the Intel results but it is interesting to see the differences in the larsonN, mstressN, and xmalloc-testN benchmarks.

Peak Working Set

The following figure shows the peak working set (rss) of the allocators on the benchmarks (on the c5.18xlarge instance).

Note that the xmalloc-testN memory usage should be disregarded as it allocates more the faster the program runs. Similarly, memory usage of mstressN, rptestN and sh8bench can vary depending on scheduling and speed. Nevertheless, even though mimalloc is fast on these benchmarks we believe the memory usage is too high and hope to improve.

References

  • [1] Emery D. Berger, Kathryn S. McKinley, Robert D. Blumofe, and Paul R. Wilson. Hoard: A Scalable Memory Allocator for Multithreaded Applications the Ninth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-IX). Cambridge, MA, November 2000. pdf

  • [2] P. Larson and M. Krishnan. Memory allocation for long-running server applications. In ISMM, Vancouver, B.C., Canada, 1998. pdf

  • [3] D. Grunwald, B. Zorn, and R. Henderson. Improving the cache locality of memory allocation. In R. Cartwright, editor, Proceedings of the Conference on Programming Language Design and Implementation, pages 177–186, New York, NY, USA, June 1993. pdf

  • [4] J. Barnes and P. Hut. A hierarchical O(n*log(n)) force-calculation algorithm. Nature, 324:446-449, 1986.

  • [5] C. Lever, and D. Boreham. Malloc() Performance in a Multithreaded Linux Environment. In USENIX Annual Technical Conference, Freenix Session. San Diego, CA. Jun. 2000. Available at https://github.com/kuszmaul/SuperMalloc/tree/master/tests

  • [6] Timothy Crundal. Reducing Active-False Sharing in TCMalloc. 2016. CS16S1 project at the Australian National University. pdf

  • [7] Alexey Kukanov, and Michael J Voss. The Foundations for Scalable Multi-Core Software in Intel Threading Building Blocks. Intel Technology Journal 11 (4). 2007

  • [8] Bobby Powers, David Tench, Emery D. Berger, and Andrew McGregor. Mesh: Compacting Memory Management for C/C++ In Proceedings of the 40th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI'19), June 2019, pages 333-–346.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.