Skip to contentSkip to conversation

AMD64 and RAID6

DISCUSSION J. Ryan Earl 3 participants 3 messages May 8, 2004 – May 11, 2004

lore
Share & tools
lore
#0 lore raw

I noticed the following in my dmesg:

raid5: measuring checksumming speed
generic_sse: 6604.000 MB/sec
raid5: using function: generic_sse (6604.000 MB/sec)
raid6: int64x1 1847 MB/s
raid6: int64x2 2753 MB/s
raid6: int64x4 2878 MB/s
raid6: int64x8 1902 MB/s
raid6: sse2x1 1015 MB/s
raid6: sse2x2 1488 MB/s
raid6: sse2x4 1867 MB/s
raid6: using algorithm sse2x4 (1867 MB/s)
md: raid6 personality registered as nr 8
md: md driver 0.90.0 MAX_MD_DEVS=256, MD_SB_DISKS=27

Why doesn't RAID6 use the int64x4 algorithm in this situation? What is the motivation of setting the 'prefer field' on the sse algorithms and not on the integer based algorithms?

From drivers/md/raid6.h:
/* Routine choices */
struct raid6_calls {
void (*gen_syndrome)(int, size_t, void **);
int (*valid)(void); /* Returns 1 if this routine set is
usable */
const char *name; /* Name of this routine set */
int prefer; /* Has special performance attribute */
};

-ryan

#1 lore raw

From: J. Ryan Earl <[email protected]>
Date: Sat, May 08, 2004 at 12:48:54PM -0500

Quoted 4 lines

Why doesn't RAID6 use the int64x4 algorithm in this situation? What is the motivation of setting the 'prefer field' on the sse algorithms and not on the integer based algorithms?

IIRC, the sse variants have better cache-behaviour, and are thus almost always selected.

Try googling for exact answers, this has come up before.

HTH,
Jurriaan

Signature · 4 lines
Spock: "Logic, logic, logic...  Logic is the beginning of wisdom,
Valeris, not  the end."
"STVI:TUC", Stardate 9522.6
Debian (Unstable) GNU/Linux 2.6.6-rc3-mm1 2x6062 bogomips 0.06 0.16
#2 lore raw
In reply to J. Ryan Earl
Followup to:  <[email protected]>
By author:    "J. Ryan Earl" <[email protected]>
In newsgroup: linux.dev.kernel
Quoted 17 lines

I noticed the following in my dmesg:

raid5: measuring checksumming speed
generic_sse: 6604.000 MB/sec
raid5: using function: generic_sse (6604.000 MB/sec)
raid6: int64x1 1847 MB/s
raid6: int64x2 2753 MB/s
raid6: int64x4 2878 MB/s
raid6: int64x8 1902 MB/s
raid6: sse2x1 1015 MB/s
raid6: sse2x2 1488 MB/s
raid6: sse2x4 1867 MB/s
raid6: using algorithm sse2x4 (1867 MB/s)
md: raid6 personality registered as nr 8
md: md driver 0.90.0 MAX_MD_DEVS=256, MD_SB_DISKS=27

Quoted 4 lines

Why doesn't RAID6 use the int64x4 algorithm in this situation? What is the motivation of setting the 'prefer field' on the sse algorithms and not on the integer based algorithms?

The SSE algorithms are non-cache-polluting. This makes them slightly slower, but avoids slowing the rest of the machine down as much.

-hpa