Skip to contentSkip to conversation

mkfs.xfs: do not set log stripe unit for probed sw <= 2

RFC PATCH NAK Eric Sandeen 1 participant 2 messages May 27, 2020

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 mkfs.xfs: do not set log stripe unit for probed sw <= 2

If the stripe width of a device is only 2x or 1x the stripe unit, there is no parity disk on this device, and setting a larger log stripe unit will not avoid any RMW cycles. However, a large log stripe unit does have significant penalties for IO amplification because every log write will be rounded up to the log stripe unit.

This was recently highlighted by a user running bonnie++ in sync mode, where the default RAID10 geometry of 256k/512k yielded results which were 4x slower than a smaller log stripe unit. While bonnie++ may not be the benchmark of choice, it does highlight this issue.

Because a larger log stripe unit yields no RMW benefit on a device with no parity disks, avoid setting in these cases.

---

... thoughts?

Am I missing a reason why we /would/ still want lsunit in this case?

mkfs/xfs_mkfs.c +4 -2
index 2e6cd280..4da69b29 100644
@@ -2407,13 +2407,15 @@ _("log stripe unit (%d) must be a multiple of the block size (%d)\n"),
 	}
 
 	/*
-	 * check that log sunit is modulo fsblksize or default it to dsunit.
+	 * check that log sunit is modulo fsblksize or default it to dsunit
+	 * if this looks like a parity device (swidth > 2x sunit).
 	 */
 	if (lsunit) {
 		/* convert from 512 byte blocks to fs blocks */
 		cfg->lsunit = DTOBT(lsunit, cfg->blocklog);
 	} else if (cfg->sb_feat.log_version == 2 &&
-		   cfg->loginternal && cfg->dsunit) {
+		   cfg->loginternal && cfg->dsunit &&
+		   (cfg->dswidth / cfg->dsunit > 2)) {
 		/* lsunit and dsunit now in fs blocks */
 		cfg->lsunit = cfg->dsunit;
 	}
author NAK
#1 lore raw
Quoted 6 lines · Eric Sandeen

If the stripe width of a device is only 2x or 1x the stripe unit, there is no parity disk on this device, and setting a larger log stripe unit will not avoid any RMW cycles. However, a large log stripe unit does have significant penalties for IO amplification because every log write will be rounded up to the log stripe unit.

Quoted 7 lines

This was recently highlighted by a user running bonnie++ in sync mode, where the default RAID10 geometry of 256k/512k yielded results which were 4x slower than a smaller log stripe unit. While bonnie++ may not be the benchmark of choice, it does highlight this issue.

Because a larger log stripe unit yields no RMW benefit on a device with no parity disks, avoid setting in these cases.

NAK

I wasn't thinking about stacked parity raids like RAID50 (thanks Dave).

So we cannot infer anything about parity disks from the top-level geometry, and so the proposed optimization is not valid.

-Eric