aboutsummaryrefslogtreecommitdiffstats
path: root/vendor/github.com/minio/md5-simd/README.md
diff options
context:
space:
mode:
Diffstat (limited to 'vendor/github.com/minio/md5-simd/README.md')
-rw-r--r--vendor/github.com/minio/md5-simd/README.md198
1 files changed, 198 insertions, 0 deletions
diff --git a/vendor/github.com/minio/md5-simd/README.md b/vendor/github.com/minio/md5-simd/README.md
new file mode 100644
index 0000000..fa6fce1
--- /dev/null
+++ b/vendor/github.com/minio/md5-simd/README.md
@@ -0,0 +1,198 @@
1
2# md5-simd
3
4This is a SIMD accelerated MD5 package, allowing up to either 8 (AVX2) or 16 (AVX512) independent MD5 sums to be calculated on a single CPU core.
5
6It was originally based on the [md5vec](https://github.com/igneous-systems/md5vec) repository by Igneous Systems, but has been made more flexible by amongst others supporting different message sizes per lane and adding AVX512.
7
8`md5-simd` integrates a similar mechanism as described in [minio/sha256-simd](https://github.com/minio/sha256-simd#support-for-avx512) for making it easy for clients to take advantages of the parallel nature of the MD5 calculation. This will result in reduced overall CPU load.
9
10It is important to understand that `md5-simd` **does not speed up** a single threaded MD5 hash sum.
11Rather it allows multiple __independent__ MD5 sums to be computed in parallel on the same CPU core,
12thereby making more efficient usage of the computing resources.
13
14## Usage
15
16[![Documentation](https://godoc.org/github.com/minio/md5-simd?status.svg)](https://pkg.go.dev/github.com/minio/md5-simd?tab=doc)
17
18
19In order to use `md5-simd`, you must first create an `Server` which can be
20used to instantiate one or more objects for MD5 hashing.
21
22These objects conform to the regular [`hash.Hash`](https://pkg.go.dev/hash?tab=doc#Hash) interface
23and as such the normal Write/Reset/Sum functionality works as expected.
24
25As an example:
26```
27 // Create server
28 server := md5simd.NewServer()
29 defer server.Close()
30
31 // Create hashing object (conforming to hash.Hash)
32 md5Hash := server.NewHash()
33 defer md5Hash.Close()
34
35 // Write one (or more) blocks
36 md5Hash.Write(block)
37
38 // Return digest
39 digest := md5Hash.Sum([]byte{})
40```
41
42To keep performance both a [Server](https://pkg.go.dev/github.com/minio/md5-simd?tab=doc#Server)
43and individual [Hasher](https://pkg.go.dev/github.com/minio/md5-simd?tab=doc#Hasher) should
44be closed using the `Close()` function when no longer needed.
45
46A Hasher can efficiently be re-used by using [`Reset()`](https://pkg.go.dev/hash?tab=doc#Hash) functionality.
47
48In case your system does not support the instructions required it will fall back to using `crypto/md5` for hashing.
49
50## Limitations
51
52As explained above `md5-simd` does not speed up an individual MD5 hash sum computation,
53unless some hierarchical tree construct is used but this will result in different outcomes.
54Running a single hash on a server results in approximately half the throughput.
55
56Instead, it allows running multiple MD5 calculations in parallel on a single CPU core.
57This can be beneficial in e.g. multi-threaded server applications where many go-routines
58are dealing with many requests and multiple MD5 calculations can be packed/scheduled for parallel execution on a single core.
59
60This will result in a lower overall CPU usage as compared to using the standard `crypto/md5`
61functionality where each MD5 hash computation will consume a single thread (core).
62
63It is best to test and measure the overall CPU usage in a representative usage scenario in your application
64to get an overall understanding of the benefits of `md5-simd` as compared to `crypto/md5`, ideally under heavy CPU load.
65
66Also note that `md5-simd` is best meant to work with large objects,
67so if your application only hashes small objects of a few kilobytes
68you may be better of by using `crypto/md5`.
69
70## Performance
71
72For the best performance writes should be a multiple of 64 bytes, ideally a multiple of 32KB.
73To help with that a [`buffered := bufio.NewWriterSize(hasher, 32<<10)`](https://golang.org/pkg/bufio/#NewWriterSize)
74can be inserted if you are unsure of the sizes of the writes.
75Remember to [flush](https://golang.org/pkg/bufio/#Writer.Flush) `buffered` before reading the hash.
76
77A single 'server' can process 16 streams concurrently with 1 core (AVX-512) or 2 cores (AVX2).
78In situations where it is likely that more than 16 streams are fully loaded it may be beneficial
79to use multiple servers.
80
81The following chart compares the multi-core performance between `crypto/md5` vs the AVX2 vs the AVX512 code:
82
83![md5-performance-overview](chart/Multi-core-MD5-Aggregated-Hashing-Performance.png)
84
85Compared to `crypto/md5`, the AVX2 version is up to 4x faster:
86
87```
88$ benchcmp crypto-md5.txt avx2.txt
89benchmark old MB/s new MB/s speedup
90BenchmarkParallel/32KB-4 2229.22 7370.50 3.31x
91BenchmarkParallel/64KB-4 2233.61 8248.46 3.69x
92BenchmarkParallel/128KB-4 2235.43 8660.74 3.87x
93BenchmarkParallel/256KB-4 2236.39 8863.87 3.96x
94BenchmarkParallel/512KB-4 2238.05 8985.39 4.01x
95BenchmarkParallel/1MB-4 2233.56 9042.62 4.05x
96BenchmarkParallel/2MB-4 2224.11 9014.46 4.05x
97BenchmarkParallel/4MB-4 2199.78 8993.61 4.09x
98BenchmarkParallel/8MB-4 2182.48 8748.22 4.01x
99```
100
101Compared to `crypto/md5`, the AVX512 is up to 8x faster (for larger block sizes):
102
103```
104$ benchcmp crypto-md5.txt avx512.txt
105benchmark old MB/s new MB/s speedup
106BenchmarkParallel/32KB-4 2229.22 11605.78 5.21x
107BenchmarkParallel/64KB-4 2233.61 14329.65 6.42x
108BenchmarkParallel/128KB-4 2235.43 16166.39 7.23x
109BenchmarkParallel/256KB-4 2236.39 15570.09 6.96x
110BenchmarkParallel/512KB-4 2238.05 16705.83 7.46x
111BenchmarkParallel/1MB-4 2233.56 16941.95 7.59x
112BenchmarkParallel/2MB-4 2224.11 17136.01 7.70x
113BenchmarkParallel/4MB-4 2199.78 17218.61 7.83x
114BenchmarkParallel/8MB-4 2182.48 17252.88 7.91x
115```
116
117These measurements were performed on AWS EC2 instance of type `c5.xlarge` equipped with a Xeon Platinum 8124M CPU at 3.0 GHz.
118
119If only one or two inputs are available the scalar calculation method will be used for the
120optimal speed in these cases.
121
122## Operation
123
124To make operation as easy as possible there is a “Server” coordinating everything. The server keeps track of individual hash states and updates them as new data comes in. This can be visualized as follows:
125
126![server-architecture](chart/server-architecture.png)
127
128The data is sent to the server from each hash input in blocks of up to 32KB per round. In our testing we found this to be the block size that yielded the best results.
129
130Whenever there is data available the server will collect data for up to 16 hashes and process all 16 lanes in parallel. This means that if 16 hashes have data available all the lanes will be filled. However since that may not be the case, the server will fill less lanes and do a round anyway. Lanes can also be partially filled if less than 32KB of data is written.
131
132![server-lanes-example](chart/server-lanes-example.png)
133
134In this example 4 lanes are fully filled and 2 lanes are partially filled. In this case the black areas will simply be masked out from the results and ignored. This is also why calculating a single hash on a server will not result in any speedup and hash writes should be a multiple of 32KB for the best performance.
135
136For AVX512 all 16 calculations will be done on a single core, on AVX2 on 2 cores if there is data for more than 8 lanes.
137So for optimal usage there should be data available for all 16 hashes. It may be perfectly reasonable to use more than 16 concurrent hashes.
138
139
140## Design & Tech
141
142md5-simd has both an AVX2 (8-lane parallel), and an AVX512 (16-lane parallel version) algorithm to accelerate the computation with the following function definitions:
143```
144//go:noescape
145func block8(state *uint32, base uintptr, bufs *int32, cache *byte, n int)
146
147//go:noescape
148func block16(state *uint32, ptrs *int64, mask uint64, n int)
149```
150
151The AVX2 version is based on the [md5vec](https://github.com/igneous-systems/md5vec) repository and is essentially unchanged except for minor (cosmetic) changes.
152
153The AVX512 version is derived from the AVX2 version but adds some further optimizations and simplifications.
154
155### Caching in upper ZMM registers
156
157The AVX2 version passes in a `cache8` block of memory (about 0.5 KB) for temporary storage of intermediate results during `ROUND1` which are subsequently used during `ROUND2` through to `ROUND4`.
158
159Since AVX512 has double the amount of registers (32 ZMM registers as compared to 16 YMM registers), it is possible to use the upper 16 ZMM registers for keeping the intermediate states on the CPU. As such, there is no need to pass in a corresponding `cache16` into the AVX512 block function.
160
161### Direct loading using 64-bit pointers
162
163The AVX2 uses the `VPGATHERDD` instruction (for YMM) to do a parallel load of 8 lanes using (8 independent) 32-bit offets. Since there is no control over how the 8 slices that are passed into the (Golang) `blockMd5` function are laid out into memory, it is not possible to derive a "base" address and corresponding offsets (all within 32-bits) for all 8 slices.
164
165As such the AVX2 version uses an interim buffer to collect the byte slices to be hashed from all 8 inut slices and passed this buffer along with (fixed) 32-bit offsets into the assembly code.
166
167For the AVX512 version this interim buffer is not needed since the AVX512 code uses a pair of `VPGATHERQD` instructions to directly dereference 64-bit pointers (from a base register address that is initialized to zero).
168
169Note that two load (gather) instructions are needed because the AVX512 version processes 16-lanes in parallel, requiring 16 times 64-bit = 1024 bits in total to be loaded. A simple `VALIGND` and `VPORD` are subsequently used to merge the lower and upper halves together into a single ZMM register (that contains 16 lanes of 32-bit DWORDS).
170
171### Masking support
172
173Due to the fact that pointers are passed directly from the Golang slices, we need to protect against NULL pointers.
174For this a 16-bit mask is passed in the AVX512 assembly code which is used during the `VPGATHERQD` instructions to mask out lanes that could otherwise result in segment violations.
175
176### Minor optimizations
177
178The `roll` macro (three instructions on AVX2) is no longer needed for AVX512 and is replaced by a single `VPROLD` instruction.
179
180Also several logical operations from the various ROUNDS of the AVX2 version could be combined into a single instruction using ternary logic (with the `VPTERMLOGD` instruction), resulting in a further simplification and speed-up.
181
182## Low level block function performance
183
184The benchmark below shows the (single thread) maximum performance of the `block()` function for AVX2 (having 8 lanes) and AVX512 (having 16 lanes). Also the baseline single-core performance from the standard `crypto/md5` package is shown for comparison.
185
186```
187BenchmarkCryptoMd5-4 687.66 MB/s 0 B/op 0 allocs/op
188BenchmarkBlock8-4 4144.80 MB/s 0 B/op 0 allocs/op
189BenchmarkBlock16-4 8228.88 MB/s 0 B/op 0 allocs/op
190```
191
192## License
193
194`md5-simd` is released under the Apache License v2.0. You can find the complete text in the file LICENSE.
195
196## Contributing
197
198Contributions are welcome, please send PRs for any enhancements. \ No newline at end of file