iar最高优化模式选择size对代码的影响_优化innochecksum实战，四倍加速榨干NVMe SSD磁盘带宽...

shell> bin/innochecksum [options] file_name

#define UNIV_PAGE_SIZE     (2 * 8192)

innodb
crc32
none

while (!feof(fil_in)) {
   bytes = read_file //read 16KB
   is_page_corrupted //using crc32/innodb/none algorithm
}

./fio --filename=test -iodepth=16 -ioengine=libaio -direct=1 -rw=write -bs=1m -size=512g -numjobs=8 -runtime=50 -group_reporting -name=test-write

write: bw=1543.2MB/s, iops=1543

./fio --filename=test -iodepth=16 -ioengine=libaio -direct=1 -rw=read -bs=2m -size=512g -numjobs=8 -runtime=50 -group_reporting -name=test-read

read : bw=2382.8MB/s, iops=2382

#include <iostream>
#include <unistd.h>
#include <cstring>
#include <sys/time.h>
#include <sys/mman.h>
#include <sys/stat.h> 
#include <vector>
#include <thread>
#include <fcntl.h>

#define WRITE_ONCE_BYTE_SIZE 4096

static char data[WRITE_ONCE_BYTE_SIZE] __attribute__((aligned(WRITE_ONCE_BYTE_SIZE))) = {'a'};
static const uint64_t kWriteCountPerThread = 1000 * 1000;
static const uint64_t kConcurrency = 64;
static const uint64_t kWriteBytesPerThread = sizeof(data) * kWriteCountPerThread;
static const uint64_t kTotalWriteBytes = kWriteBytesPerThread * kConcurrency;

uint64_t NowMicros() {
  struct timeval tv;
  gettimeofday(&tv, NULL);
  return static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
}

void writer(int index) {
  std::string fname = "data" + std::to_string(index);

  int fd = ::open(fname.c_str(), O_NOATIME | O_RDWR | O_CREAT, 0644);

  int ret = posix_fallocate(fd, 0, kWriteBytesPerThread);
  if (ret != 0) { 
    printf("fallocate err %dn", ret);
  }
  lseek(fd, 0, SEEK_SET);

  for (int32_t i = 0; i < kWriteCountPerThread; i++) {
    ::write(fd, data, WRITE_ONCE_BYTE_SIZE);
  }

  close(fd);
}

int main() {
  uint64_t st, ed;
  st = NowMicros();
  std::vector<std::thread> threads;
  for(int i = 0; i < kConcurrency; i++) {
    std::thread worker(writer, i);
    threads.push_back(std::move(worker));
  }
  for (int i = 0; i < kConcurrency; i++) {
    threads[i].join();
  }
  ed = NowMicros();
  printf("time elapsed microsecond(us) %lld, %lld MB/sn", ed - st, kTotalWriteBytes / (ed - st));
  return 0;
}

procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu-----
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
65  0      0 3618176 330256 247338112    0    0     0 1457344 152720 6513  0 96  4  0  0
65  0      0 3492712 330256 247428272    0    0     0 1315880 169502 4963  0 96  4  0  0
66  0      0 3545536 330256 247339136    0    0     0 1183212 180769 4215  0 96  4  0  0
65  0      0 3525324 330256 247325776    0    0     0 1309008 183833 4679  0 97  3  0  0
66  0      0 3590052 330256 247226944    0    0     0 1496924 176723 5951  0 97  3  0  0
57  0      0 3543704 330256 247167072    0    0     0 2441148 197181 7717  0 95  5  0  0
64  0      0 3565740 330256 247107824    0    0     0 1472412 170361 8038  0 94  6  0  0
55  0      0 3520116 330256 247132928    0    0     0 1457048 154563 9472  0 95  5  0  0
65  1      0 3549728 330256 247052544    0    0     0 1555564 133119 8827  0 92  8  0  0
63  0      0 3567856 330256 247002144    0    0     0 1343976 144924 7134  0 97  3  0  0

time elapsed microsecond(us) 123703672, 1589 MB/s

char* base = nullptr;
char* cursor = nullptr;
uint64_t staging_offset = 0, file_offset = 0;

int staging_fd = ::open(staging_fname.c_str(), O_NOATIME | O_RDWR | O_CREAT, 0644);
int ret = posix_fallocate(staging_fd, 0, kStagingFileSize);
void* staging_ptr = mmap(NULL, kStagingFileSize, PROT_READ | PROT_WRITE,
        MAP_SHARED, staging_fd, 0);

int fd = ::open(fname.c_str(), O_DIRECT | O_NOATIME | O_RDWR | O_CREAT, 0644);
ret = posix_fallocate(fd, 0, kWriteBytesPerThread);
lseek(fd, 0, SEEK_SET);

base = static_cast<char*>(staging_ptr);
cursor = static_cast<char*>(staging_ptr);
for (int32_t i = 0; i < kWriteCountPerThread; i++) {
    if (unlikely(staging_offset >= kStagingFileSize)) {
        pwrite64(fd, staging_ptr, kStagingFileSize, file_offset);
        cursor = base;
        file_offset += kStagingFileSize;
        staging_offset = 0;
    }
    memcpy(cursor, data, 4096);
    cursor += 4096;
    staging_offset += 4096;
}

close(staging_fd);
close(fd);
munmap(staging_ptr, kStagingFileSize);
staging_ptr = nullptr;

time elapsed microsecond(us) 170181550, 1540 MB/s

int fd = ::open(fname.c_str(), O_NOATIME | O_RDWR, 0644);
void* buffer = NULL;
posix_memalign(&buffer, getpagesize(), kReadOnceByteSize);
uint64_t offset = 0;
for (int32_t i = 0; i < kReaderCountPerThread; i++) {
    pread64(fd, buffer, kReadOnceByteSize, offset);
    offset += kReadOnceByteSize;
}
close(fd);

time elapsed microsecond(us) 109012309, 2404 MB/s

time elapsed microsecond(us) 90598873, 2531 MB/s

$ls
Makefile
fast_innochecksum.cpp
partition.cpp
partition.h
partition_scan_cache_pool.cpp
partition_scan_cache_pool.h
scanner.cpp
scanner.h
shard.cpp
shard.h
thread_pool.h
ut0crc32.cpp
ut0crc32.h
util.cpp
util.h

+-------------------------------------------+
|IO    |thread|Shard   |Partition|Throughput|
|mode  |count |ByteSize|ByteSize |          |
+-------------------------------------------+
|direct|   1  |  256MB |   16MB  | 2081 MB/s|
|direct|   1  |    1G  |   64MB  | 2173 MB/s|
|direct|  16  |  256MB |   16MB  | 2468 MB/s|
|direct|  16  |    1G  |   64MB  | 2503 MB/s|
|buffer|   1  |  256MB |   16MB  | 1086 MB/s|
|buffer|   1  |    1G  |   64MB  | 1109 MB/s|
|buffer|  16  |  256MB |   16MB  | 2342 MB/s|
|buffer|  16  |    1G  |   64MB  | 2386 MB/s|
+-------------------------------------------+
Note: kNumOfScanCachePartition=3