jasder
/
simdjson
1
0
Fork 0
Code Issues Pull Requests Releases 2 Wiki Activity

[no ci] update to power usage section

This commit is contained in:
Daniel Lemire 2021-07-27 10:30:19 -04:00 committed by GitHub
parent eb93b98d6a
commit e681234e45
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
doc/performance.md
View File

@ -10,7 +10,7 @@ are still some scenarios where tuning can enhance performance.
* [Large files and huge page support](#large-files-and-huge-page-support)
* [Number parsing](#number-parsing)
* [Visual Studio](#visual-studio)
* [Downclocking](#downclocking)
* [Power Usage and Downclocking](#power-usage-and-downclocking)
Reusing the parser for maximum efficiency
-----------------------------------------
@ -110,19 +110,15 @@ Recent versions of Microsoft Visual Studio on Windows provides support for the L
Under Windows, we also support the GNU GCC compiler via MSYS2. The performance of 64-bit MSYS2 under Windows excellent (on par with Linux).
Downclocking
Power Usage and Downclocking
--------------
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 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](https://golang.org/src/runtime/memmove_amd64.s), [Glibc](https://github.com/ihtsae/glibc/commit/5f3d0b78e011d2a72f9e88b0e9ef5bc081d18f97), [LLVM](https://github.com/llvm/llvm-project/blob/96f3ea0d21b48ca088355db10d4d1a2e9bc9f884/lldb/tools/debugserver/source/MacOSX/i386/DNBArchImplI386.cpp), [Rust](https://github.com/rust-lang/rust/commit/070fad1701fb36b112853b0a6a9787a7bb7ff34c), [Java](http://hg.openjdk.java.net/jdk8u/jdk8u/hotspot/file/c1374141598c/src/cpu/x86/vm/stubGenerator_x86_64.cpp#l1297), [PHP](https://github.com/php/php-src/blob/e5cb53ec68603d4dbdd780fd3ecfca943b4fd383/ext/standard/string.c)). 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:
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](https://lemire.me/blog/2018/09/07/avx-512-when-and-how-to-use-these-new-instructions/).
- 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).