From: Jeff Lien
> Sent: 10 August 2018 20:12
>
> This patch provides a performance improvement for the CRC16 calculations done in read/write
> workloads using the T10 Type 1/2/3 guard field. For example, today with sequential write
> workloads (one thread/CPU of IO) we consume 100% of the CPU because of the CRC16 computation
> bottleneck. Today's block devices are considerably faster, but the CRC16 calculation prevents
> folks from utilizing the throughput of such devices. To speed up this calculation and expose
> the block device throughput, we slice the old single byte for loop into a 16 byte for loop,
> with a larger CRC table to match. The result has shown 5x performance improvements on various
> big endian and little endian systems running the 4.18.0 kernel version.
>
> FIO Sequential Write, 64K Block Size, Queue Depth 64
> BE Base Kernel: bw=201.5 MiB/s
> BE Modified CRC Calc: bw=968.1 MiB/s
> 4.80x performance improvement
>
> LE Base Kernel: bw=357 MiB/s
> LE Modified CRC Calc: bw=1964 MiB/s
> 5.51x performance improvement
>
> FIO Sequential Read, 64K Block Size, Queue Depth 64
> BE Base Kernel: bw=611.2 MiB/s
> BE Modified CRC calc: bw=684.9 MiB/s
> 1.12x performance improvement
>
> LE Base Kernel: bw=797 MiB/s
> LE Modified CRC Calc: bw=2730 MiB/s
> 3.42x performance improvement
>
> Reviewed-by: Dave Darrington <david.darrington@xxxxxxx>
> Reviewed-by: Jeff Furlong <jeff.furlong@xxxxxxx>
> Signed-off-by: Jeff Lien <jeff.lien@xxxxxxx>
> ---
> crypto/crct10dif_common.c | 605 +++++++++++++++++++++++++++++++++++++++++++---
> 1 file changed, 569 insertions(+), 36 deletions(-)
>
> diff --git a/crypto/crct10dif_common.c b/crypto/crct10dif_common.c
> index b2fab36..40e1d6c 100644
> --- a/crypto/crct10dif_common.c
> +++ b/crypto/crct10dif_common.c
> @@ -32,47 +32,580 @@
> * x^16 + x^15 + x^11 + x^9 + x^8 + x^7 + x^5 + x^4 + x^2 + x + 1
> * gt: 0x8bb7
> */
> -static const __u16 t10_dif_crc_table[256] = {
...
> + }
> };
>
> __u16 crc_t10dif_generic(__u16 crc, const unsigned char *buffer, size_t len)
> {
> - unsigned int i;
> + const __u8 *i = (const __u8 *)buffer;
> + const __u8 *i_end = i + len;
> + const __u8 *i_last16 = i + (len / 16 * 16);
>
> - for (i = 0 ; i < len ; i++)
> - crc = (crc << 8) ^ t10_dif_crc_table[((crc >> 8) ^ buffer[i]) & 0xff];
> + for (; i < i_last16; i += 16) {
> + crc = t10_dif_crc_table[15][i[0] ^ (__u8)(crc >> 8)] ^
> + t10_dif_crc_table[14][i[1] ^ (__u8)(crc >> 0)] ^
> + t10_dif_crc_table[13][i[2]] ^
> + t10_dif_crc_table[12][i[3]] ^
> + t10_dif_crc_table[11][i[4]] ^
> + t10_dif_crc_table[10][i[5]] ^
> + t10_dif_crc_table[9][i[6]] ^
> + t10_dif_crc_table[8][i[7]] ^
> + t10_dif_crc_table[7][i[8]] ^
> + t10_dif_crc_table[6][i[9]] ^
> + t10_dif_crc_table[5][i[10]] ^
> + t10_dif_crc_table[4][i[11]] ^
> + t10_dif_crc_table[3][i[12]] ^
> + t10_dif_crc_table[2][i[13]] ^
> + t10_dif_crc_table[1][i[14]] ^
> + t10_dif_crc_table[0][i[15]];
...
I suspect that all the gain comes from a slight relaxation of the
register dependency chain and the loop unrolling.
A more interesting version would be to generate the lookup table
for a byte followed by 3 zero bytes.
You could then run four separate register dependency chains using the
same 256 entry lookup table.
A little bit of work at the end of the buffer should sort it all out.
There is also the lookup table free version:
uint32_t
crc_step(uint32_t crc, uint32_t byte_val)
{
uint32_t t = crc ^ (byte_val & 0xff);
t = (t ^ t << 4) & 0xff;
return crc >> 8 ^ t << 8 ^ t << 3 ^ t >> 4;
}
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
