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.

It is a drop-in replacement for malloc and can be used in other programs without code changes, for example, on Unix you can use it as:

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

Notable aspects of the design include:

• small and consistent: the library is less than 3500 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 64KB 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.
• lazy initialization: pages in a segment are lazily initialized so no memory is touched until it becomes allocated, reducing the resident memory and potential page faults.
• secure: mimalloc can be build in secure mode, adding guard pages, randomized allocation, encoded free lists, etc. to protect against various heap vulnerabilities. The performance penalty is only around 3% 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 16.7% waste in allocation sizes), and has no internal points of contention using only atomic operations.
• fast: In our benchmarks (see below), mimalloc always outperforms all 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.

You can read more on the design of mimalloc in the upcoming technical report.

Enjoy!

Building

Windows

Open ide/vs2017/mimalloc.sln in Visual Studio 2017 and build. 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.

MacOSX, Linux, BSD, etc.

We use cmake¹ as the build system:

• cd out/release

• cmake ../.. (generate the make file)

• make (and build)

  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:

• cd out/debug

• cmake -DCMAKE_BUILD_TYPE=Debug ../..

• make

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

Or build with clang:

• CC=clang cmake ../..

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 1.0 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.

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

> env MIMALLOC_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.

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.

Unix, BSD, MacOSX

On these 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 (on Linux, BSD, etc.)

• env DYLD_INSERT_LIBRARIES=usr/lib/libmimalloc.dylib myprogram (On MacOSX)

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

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_STATS=1 LD_PRELOAD=/usr/lib/libmimalloc-debug.so myprogram

Windows

On Windows you need to link your program explicitly with the mimalloc DLL, and use the C-runtime library as a DLL (the /MD or /MDd switch). To ensure the mimalloc DLL gets loaded it is easiest to insert some call to the mimalloc API in the main function, like mi_version().

Due to the way mimalloc intercepts the standard malloc at runtime, it is best to link to the mimalloc import library first on the command line so it gets loaded right after the universal C runtime DLL (ucrtbase). See the mimalloc-override-test project for an example.

Static override

On Unix 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 (or .obj)). 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 ...

Performance

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.

Allocators are interesting as there exists no algorithm that is generally 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 underperformance in less common situations (which is what the second half of our benchmark set tests for).

In our benchmarks, mimalloc always outperforms all 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 the wide range of benchmarks.

The benchmark suite is scripted and available separately as mimalloc-bench.

Tested Allocators

We tested mimalloc with 9 leading allocators over 12 benchmarks and the SpecMark benchmarks. The tested allocators are:

• mi: The mimalloc allocator, using version tag v1.0.0. We also test a secure version of mimalloc as smi which uses the techniques described in Section [#sec-secure].
• tc: The tcmalloc allocator which comes as part of the Google performance tools and is used in the Chrome browser. Installed as package libgoogle-perftools-dev version 2.5-2.2ubuntu3.
• je: The jemalloc allocator by Jason Evans is developed at Facebook and widely used in practice, for example in FreeBSD and Firefox. Using version tag 5.2.0.
• sn: The snmalloc allocator is a recent concurrent message passing allocator by Liétar et al. [8]. Using git-0b64536b.
• rp: The rpmalloc allocator uses 32-byte aligned allocations and is developed by Mattias Jansson at Rampant Pixels. Using version tag 1.3.1.
• hd: The Hoard allocator by Emery Berger [1]. This is one of the first multi-thread scalable allocators. Using version tag 3.13.
• glibc: The system allocator. Here we use the glibc allocator (which is originally based on Ptmalloc2), using version 2.27.0. Note that version 2.26 significantly improved scalability over earlier versions.
• sm: The Supermalloc allocator by Bradley Kuszmaul uses hardware transactional memory to speed up parallel operations. Using version git-709663fb.
• tbb: The Intel TBB allocator that comes with the Thread Building Blocks (TBB) library [7]. Installed as package libtbb-dev, version 2017~U7-8.

All allocators run exactly the same benchmark programs on Ubuntu 18.04.1 and use LD_PRELOAD to override the default allocator. The wall-clock elapsed time and peak resident memory (rss) are measured with the time program. The average scores over 5 runs are used. Performance is reported relative to mimalloc, e.g. a time of 1.5× means that the program took 1.5× longer than mimalloc.

Benchmarks

The first set of benchmarks are real world programs and consist of:

• cfrac: by Dave Barrett, implementation of continued fraction factorization which uses many small short-lived allocations -- exactly the workload we are targeting for Koka and Lean.
• espresso: a programmable logic array analyzer, described by Grunwald, Zorn, and Henderson [3]. in the context of cache aware memory allocation.
• barnes: a hierarchical n-body particle solver [4] which uses relatively few allocations compared to cfrac and espresso. Simulates the gravitational forces between 163840 particles.
• leanN: The Lean compiler by de Moura et al, version 3.4.1, compiling its own standard library concurrently using N threads (./lean --make -j N). Big real-world workload with intensive allocation.
• redis: running the redis 5.0.3 server on 1 million requests pushing 10 new list elements and then requesting the head 10 elements. Measures the requests handled per second.
• larsonN: by Larson and Krishnan [2]. Simulates a server workload using 100 separate threads which each allocate and free many objects but leave some objects to be freed by other threads. Larson and Krishnan observe this behavior (which they call bleeding) in actual server applications, and the benchmark simulates this.

The second set of benchmarks are stress tests and consist of:

• alloc-test: a modern allocator test developed by OLogN Technologies AG (ITHare.com) Simulates intensive allocation workloads with a Pareto size distribution. The alloc-testN benchmark runs on N cores doing 100·10^6^ allocations per thread with objects up to 1KiB in size. Using commit 94f6cb (master, 2018-07-04)
• sh6bench: by MicroQuill as part of SmartHeap. Stress test where some of the objects are freed in a usual last-allocated, first-freed (LIFO) order, but others are freed in reverse order. Using the public source (retrieved 2019-01-02)
• sh8benchN: by MicroQuill as part of SmartHeap. Stress test for multi-threaded allocation (with N threads) where, just as in larson, some objects are freed by other threads, and some objects freed in reverse (as in sh6bench). Using the public source (retrieved 2019-01-02)
• xmalloc-testN: by Lever and Boreham [5] and Christian Eder. We use the updated version from the SuperMalloc repository. This is a more extreme version of the larson benchmark with 100 purely allocating threads, and 100 purely deallocating threads with objects of various sizes migrating between them. This asymmetric producer/consumer pattern is usually difficult to handle by allocators with thread-local caches.
• cache-scratch: by Emery Berger [1]. Introduced with the Hoard allocator to test for passive-false sharing of cache lines: first some small objects are allocated and given to each thread; the threads free that object and allocate immediately another one, and access that repeatedly. If an allocator allocates objects from different threads close to each other this will lead to cache-line contention.

On a 16-core AMD EPYC running Linux

Testing on a big Amazon EC2 instance (r5a.4xlarge) consisting of a 16-core AMD EPYC 7000 at 2.5GHz with 128GB ECC memory, running Ubuntu 18.04.1 with LibC 2.27 and GCC 7.3.0. We excluded SuperMalloc here as it use transactional memory instructions that are usually not supported in a virtualized environment.

Memory usage:

(note: the xmalloc-testN memory usage should be disregarded is it allocates more the faster the program runs).

In the first five benchmarks we can see mimalloc outperforms the other allocators moderately, but we also see that all these modern allocators perform well -- the times of large performance differences in regular workloads are over. In cfrac and espresso, mimalloc is a tad faster than tcmalloc and jemalloc, but a solid 10% faster than all other allocators on espresso. The tbb allocator does not do so well here and lags more than 20% behind mimalloc. The cfrac and espresso programs do not use much memory (~1.5MB) so it does not matter too much, but still mimalloc uses about half the resident memory of tcmalloc.

The leanN program is most interesting as a large realistic and concurrent workload and there is a 8% 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 redis benchmark shows more differences between the allocators where mimalloc is 14% faster than jemalloc. On this benchmark tbb (and Hoard) do not do well and are over 40% slower.

The larson server workload which allocates and frees objects between many threads shows even larger differences, where mimalloc is more than 2.5× faster than tcmalloc and jemalloc which is quite surprising for these battle tested allocators -- probably due to the object migration between different threads. This is a difficult benchmark for other allocators too where mimalloc is still 48% faster than the next fastest (snmalloc).

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

The alloc-test is very allocation intensive 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, snmalloc, and Hoard seem to scale less well and do more than 10% worse on the multi-core version. Even the best allocators (tcmalloc and jemalloc) are more than 10% slower as mimalloc here.

Also in sh6bench mimalloc does much better than the others (more than 2× faster than jemalloc). We cannot explain this well but believe it is caused in part by the "reverse" free-ing pattern in sh6bench.

Again in sh8bench the mimalloc allocator handles object migration between threads much better and is over 36% faster than the next best allocator, snmalloc. Whereas tcmalloc did well on sh6bench, the addition of object migration caused it to be almost 3 times slower than before.

The xmalloc-testN benchmark simulates an asymmetric workload where some threads only allocate, and others only free. The snmalloc allocator was especially developed to handle this case well as it often occurs in concurrent message passing systems. Here we see that the mimalloc technique of having non-contended sharded thread free lists pays off and it even outperforms snmalloc. Only jemalloc also handles this reasonably well, while the others underperform by a large margin. The optimization on mimalloc to do a delayed free only once for full pages is quite important -- without it mimalloc is almost twice as slow (as then all frees contend again on the single heap delayed free list).

The cache-scratch benchmark also demonstrates the different architectures of the allocators nicely. With a single thread they all perform the same, but when running with multiple threads the allocator induced false sharing of the cache lines causes large run-time differences, where mimalloc is more than 18× faster than jemalloc and tcmalloc! 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 snmalloc and tbb also avoid the cache line sharing like mimalloc. Kukanov and Voss [7] describe in detail how the design of tbb avoids the false cache line sharing. The Hoard allocator is also specifically designed to avoid this false sharing and we are not sure why it is not doing well here (although it runs still 5× as fast as tcmalloc).

On a 4-core Intel Xeon workstation

Below are the benchmark results on an HP Z4-G4 workstation with a 4-core Intel® Xeon® W2123 at 3.6 GHz with 16GB ECC memory, running Ubuntu 18.04.1 with LibC 2.27 and GCC 7.3.0.

Memory usage:

(note: the xmalloc-testN memory usage should be disregarded is it allocates more the faster the program runs).

This time SuperMalloc (sm) is included as this platform supports hardware transactional memory. Unfortunately, there are no entries for SuperMalloc in the leanN and xmalloc-testN benchmarks as it faulted on those. We also added the secure version of mimalloc as smi.

Overall, the relative results are quite similar as before. Most allocators fare better on the larsonN benchmark now -- either due to architectural changes (AMD vs. Intel) or because there is just less concurrency. Unfortunately, the SuperMalloc faulted on the leanN and xmalloc-testN benchmarks.

The secure mimalloc version uses guard pages around each (mimalloc) page, encodes the free lists and uses randomized initial free lists, and we expected it would perform quite a bit worse -- but on the first benchmark set it performed only about 3% slower on average, and is second best overall.

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. http://​courses.​cecs.​anu.​edu.​au/​courses/​CSPROJECTS/​16S1/​Reports/​Timothy*​Crundal*​Report.​pdf. CS16S1 project at the Australian National University.

• [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] Paul Liétar, Theodore Butler, Sylvan Clebsch, Sophia Drossopoulou, Juliana Franco, Matthew J Parkinson, Alex Shamis, Christoph M Wintersteiger, and David Chisnall. Snmalloc: A Message Passing Allocator. In Proceedings of the 2019 ACM SIGPLAN International Symposium on Memory Management, 122–135. ACM. 2019.