Skip to content

Latest commit

 

History

History
167 lines (122 loc) · 11.1 KB

performance.md

File metadata and controls

167 lines (122 loc) · 11.1 KB
Raw

Performance Notes

simdjson strives to be at its fastest without tuning, and generally achieves this. However, there are still some scenarios where tuning can enhance performance.

Reusing the parser for maximum efficiency

If you're using simdjson to parse multiple documents, or in a loop, you should make a parser once and reuse it. The simdjson library will allocate and retain internal buffers between parses, keeping buffers hot in cache and keeping memory allocation and initialization to a minimum. In this manner, you can parse terabytes of JSON data without doing any new allocation.

ondemand::parser parser;

// This initializes buffers  big enough to handle this JSON.
auto json = "[ true, false ]"_padded;
auto doc = parser.iterate(json);
for(bool i : doc.get_array()) {
  cout << i << endl;
}

// This reuses the existing buffers
auto number_json = "[1, 2, 3]"_padded;
doc = parser.iterate(number_json);
for(int64_t i : doc.get_array()) {
  cout << i << endl;
}

Reusing string buffers

We recommend against creating many std::string or simdjson::padded_string instances to store the JSON content in your application. Creating many non-trivial objects is convenient but often surprisingly slow. Instead, as much as possible, you should allocate (once or a few times) reusable memory buffers where you write your JSON content. If you have a buffer json_str (of type char*) allocated for capacity bytes and you store a JSON document spanning length bytes, you can pass it to simdjson as follows:

 auto doc = parser.iterate(padded_string_view(json_str, length, capacity));

or simply

 auto doc = parser.iterate(json_str, length, capacity);

Server Loops: Long-Running Processes and Memory Capacity

The On Demand approach also automatically expands its memory capacity when larger documents are parsed. However, for longer processes where very large files are processed (such as server loops), this capacity is not resized down. On Demand also lets you adjust the maximal capacity that the parser can process:

  • You can set an upper bound (max_capacity) when construction the parser:
    ondemand::parser parser(1000*1000);  // Never grows past documents > 1 MB
    auto doc = parser.iterate(json);
    for (web_request request : listen()) {
      padded_string json;
      padded_string json = padded_string::load(request.body);
      auto error = parser.iterate(json);
      // If the document was above our limit, emit 413 = payload too large
      if (error == CAPACITY) { request.respond(413); continue; }
      // ...
    }

The capacity will grow as the parser encounters larger documents up to 1 MB.

  • You can also allocate a fixed capacity that will never grow:
    ondemand::parser parser(1000*1000);
    parser.allocate(1000*1000)  // Fix the capacity to 1 MB
    auto doc = parser.iterate(json);
    for (web_request request : listen()) {
      padded_string json;
      padded_string json = padded_string::load(request.body);
      auto error = parser.iterate(json);
      // If the document was above our limit, emit 413 = payload too large
      if (error == CAPACITY) { request.respond(413); continue; }
      // ...
    }

You can also manually set the maximal capacity using the method set_max_capacity().

Large files and huge page support

There is a memory allocation performance cost the first time you process a large file (e.g. 100MB). Between the cost of allocation, the fact that the memory is not in cache, and the initial zeroing of memory, on some systems, allocation runs far slower than parsing (e.g., 1.4GB/s). Reusing the parser mitigates this by paying the cost once, but does not eliminate it.

In large file use cases, enabling transparent huge page allocation on the OS can help a lot. We haven't found the right way to do this on Windows or OS/X, but on Linux, you can enable transparent huge page allocation with a command like:

echo always > /sys/kernel/mm/transparent_hugepage/enabled

In general, when running benchmarks over large files, we recommend that you report performance numbers with and without huge pages if possible. Furthermore, you should amortize the parsing (e.g., by parsing several large files) to distinguish the time spent parsing from the time spent allocating memory. If you are using the parse benchmarking tool provided with the simdjson library, you can use the -H flag to omit the memory allocation cost from the benchmark results.

./parse largefile # includes memory allocation cost
./parse -H largefile # without memory allocation

Number parsing

Some JSON files contain many floating-point values. It is the case with many GeoJSON files. Accurately parsing decimal strings into binary floating-point values with proper rounding is challenging. To our knowledge, it is not possible, in general, to parse streams of numbers at gigabytes per second using a single core. While using the simdjson library, it is possible that you might be limited to a few hundred megabytes per second if your JSON documents are densely packed with floating-point values.

  • When possible, you should favor integer values written without a decimal point, as it simpler and faster to parse decimal integer values.
  • When serializing numbers, you should not use more digits than necessary: 17 digits is all that is needed to exactly represent double-precision floating-point numbers. Using many more digits than necessary will make your files larger and slower to parse.
  • When benchmarking parsing speeds, always report whether your JSON documents are made mostly of floating-point numbers when it is the case, since number parsing can then dominate the parsing time.

Visual Studio

On Intel and AMD Windows platforms, Microsoft Visual Studio enables programmers to build either 32-bit (x86) or 64-bit (x64) binaries. We urge you to always use 64-bit mode. Visual Studio 2019 should default on 64-bit builds when you have a 64-bit version of Windows, which we recommend.

When compiling with Visual Studio, we recommend the flags /Ob2 /O2 or better. We do not recommend that you compile simdjson with architecture-specific flags such as arch:AVX2. The simdjson library automatically selects the best execution kernel at runtime.

Recent versions of Microsoft Visual Studio on Windows provides support for the LLVM Clang compiler. You only need to install the "Clang compiler" optional component (ClangCL). You may also get a copy of the 64-bit LLVM CLang compiler for Windows directly from LLVM. The simdjson library fully supports the LLVM Clang compiler under Windows. In fact, you may get better performance out of simdjson with the LLVM Clang compiler than with the regular Visual Studio compiler. Meanwhile the LLVM CLang compiler is binary compatible with Visual Studio which means that you can combine their binaries (executables and libraries).

Under Windows, we also support the GNU GCC compiler via MSYS2. The performance of 64-bit MSYS2 under Windows excellent (on par with Linux).

Power Usage and Downclocking

The simdjson library relies on SIMD instructions. SIMD instructions are the public transportation of computing. Instead of using 4 distinct instructions to add numbers, you can replace them with a single instruction that does the same work. Though the one instruction is slightly more expensive, the energy used per unit of work is much less with SIMD. If you can increase your speed using SIMD instructions (NEON, SSE, AVX), you should expect to reduce your power usage.

The SIMD instructions that simdjson relies upon (SSE and AVX under x64, NEON under ARM, ALTIVEC under PPC) are routinely part of runtime libraries (e.g., Go, Glibc, LLVM, Rust, Java, PHP). What distinguishes the simdjson library is that it is built from the ground up to benefit from these instructions.

You should not expect the simdjson library to cause downclocking of your recent Intel CPU cores. On some Intel processors, using SIMD instructions in a sustained manner on the same CPU core may result in a phenomenon called downclocking whereas the processor initially runs these instructions at a slow speed before reducing the frequency of the core for a short time (milliseconds). Intel refers to these states as licenses. On some current Intel processors, it occurs under two scenarios:

  • Whenever 512-bit AVX-512 instructions are used.
  • Whenever heavy 256-bit or wider instructions are used. Heavy instructions are those involving floating point operations or integer multiplications (since these execute on the floating point unit).

The simdjson library does not make use of heavy 256-bit instructions. We do use vectorized multiplications, but only using 128-bit registers. Thus there should be no downclocking due to simdjson on recent processors, except when AVX-512 is allowed and detected. However, we only allow AVX-512 on recent processors (Ice Lake/Tiger Lake or better) where little to no frequency throttling is expected. If you can still concerned, you can easily disable AVX-512 with the CMake option SIMDJSON_AVX512_ALLOWED set to OFF (e.g., cmake -D SIMDJSON_AVX512_ALLOWED=OFF -B build && cmake --build build) or by setting the macro SIMDJSON_AVX512_ALLOWED to 0 in C++ prior to importing the headers.

You may still be worried about which SIMD instruction set is used by simdjson. Thankfully, you can always determine and change which architecture-specific implementation is used by simdjson. Thus even if your CPU supports AVX2, you do not need to use AVX2. You are in control.