Skip to contentSkip to conversation

Add maxthinktime to sysfs

PATCH In review FabF 3 participants 5 messages Apr 30, 2004 – May 1, 2004 Block layer

lore
Share & tools
lore

FabF

author
#0 lore raw

PATCH 1/1 Add maxthinktime to sysfs

Andrew,

Here's a patch to add the asio maxthinktime to sysfs.
Could you apply ?

Regards,
Fabian

drivers/block/as-iosched.c +9 -2
diff -Naur orig/drivers/block/as-iosched.c edited/drivers/block/as-iosched.c
@@ -65,7 +65,7 @@
  * or doing a lengthy computation. A small penalty can be justified there, and
  * will still catch out those processes that constantly have large thinktimes.
  */
-#define MAX_THINKTIME (HZ/50UL)
+unsigned long maxthinktime=(HZ/50UL);
 
 /* Bits in as_io_context.state */
 enum as_io_states {
@@ -869,7 +869,7 @@
 			if (test_bit(AS_TASK_IORUNNING, &aic->state)
 							&& in_flight == 0) {
 				thinktime = jiffies - aic->last_end_request;
-				thinktime = min(thinktime, MAX_THINKTIME-1);
+				thinktime = min(thinktime, maxthinktime-1);
 			} else
 				thinktime = 0;
 			as_update_thinktime(ad, aic, thinktime);
@@ -1951,6 +1951,7 @@
 {									\
 	return as_var_show(__VAR, (page));			\
 }
+SHOW_FUNCTION(as_maxthinktime_show, maxthinktime);
 SHOW_FUNCTION(as_readexpire_show, ad->fifo_expire[REQ_SYNC]);
 SHOW_FUNCTION(as_writeexpire_show, ad->fifo_expire[REQ_ASYNC]);
 SHOW_FUNCTION(as_anticexpire_show, ad->antic_expire);
@@ -1968,6 +1969,7 @@
 		*(__PTR) = (MAX);					\
 	return ret;							\
 }
+STORE_FUNCTION(as_maxthinktime_store, &maxthinktime, 0, LONG_MAX);
 STORE_FUNCTION(as_readexpire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
 STORE_FUNCTION(as_writeexpire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
 STORE_FUNCTION(as_anticexpire_store, &ad->antic_expire, 0, INT_MAX);
@@ -1977,6 +1979,11 @@
 			&ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
 #undef STORE_FUNCTION
 
+static struct as_fs_entry as_maxthinktime = {
+	.attr = {.name = "maxthinktime", .mode = S_IRUGO | S_IWUSR },
+	.show = as_maxthinktime_show,
+	.store = as_maxthinktime_store,
+};
 static struct as_fs_entry as_est_entry = {
 	.attr = {.name = "est_time", .mode = S_IRUGO },
 	.show = as_est_show,
#1 lore raw
Quoted 2 lines · FabF

Here's a patch to add the asio maxthinktime to sysfs.

Why? Have you measured any benefit from varying it, and if so, what was the result?

Badari, did you find any need to vary this in the AS tuning work which you were doing? (What happened to that, btw?)

If we're going to expose this tunable to users it needs to be documented in as-iosched.txt

author
#2 lore raw
Quoted 6 lines · Andrew Morton
Quoted 2 lines · FabF

Here's a patch to add the asio maxthinktime to sysfs.

Why? Have you measured any benefit from varying it, and if so, what was the result?

Andrew,

I notice huge variations in first seconds of a 10 client
throughput activity as attached (5,100,300 as maxthinktime). It's just another parameter I'd like to play with ;)
Maybe this one's exposition is not important intrinsequely but we could bring interesting combinations later ...

Quoted 6 lines

Badari, did you find any need to vary this in the AS tuning work which you were doing? (What happened to that, btw?)

If we're going to expose this tunable to users it needs to be documented in as-iosched.txt

I'm an absolute beginner Andrew ... Maybe someone more experienced could do that ?

PS:Attached the patch v2.

Best regards,
Fabian

ffbench 11 - Fabian Frederick 04/2004 - asio - maxthinktime - 10 clients
5 100 300
---------------------
5.37 21.26 16.28
11.71 23.10 20.14
15.22 21.05 20.20
13.85 18.03 17.55
11.58 16.39 16.86
11.41 15.50 14.30
11.43 14.37 14.03
11.43 13.82 14.16
11.43 13.60 14.29
11.57 13.27 14.01
11.24 13.06 13.67
10.78 12.91 13.67
10.71 12.41 12.91
10.91 12.53 12.70
10.97 12.32 12.67
10.90 12.34 12.82
10.62 12.25 12.68
10.59 11.81 12.61
10.44 11.40 12.15
10.44 11.35 11.96
10.37 11.36 11.68
10.20 11.19 11.73
10.39 11.18 11.85
10.22 11.22 11.92
10.10 11.15 11.81
10.11 11.03 11.53
9.95 11.00 11.56
10.15 11.21 11.55
10.32 11.17 11.54
10.34 10.96 11.47
10.08 11.02 11.47
10.10 11.14 11.54
10.15 11.14 11.45
10.07 11.04 11.31
10.12 11.08 11.26
10.01 11.04 11.28
10.02 10.93 11.15
9.98 10.70 10.93
9.94 10.60 10.93
9.85 10.45 10.78
9.87 10.40 10.68
9.98 10.37 10.62
9.95 10.28 10.56
9.87 10.20 10.51
9.72 10.16 10.55

drivers/block/as-iosched.c +10 -2
diff -Naur orig/drivers/block/as-iosched.c edited/drivers/block/as-iosched.c
@@ -65,7 +65,7 @@
  * or doing a lengthy computation. A small penalty can be justified there, and
  * will still catch out those processes that constantly have large thinktimes.
  */
-#define MAX_THINKTIME (HZ/50UL)
+unsigned long maxthinktime=(HZ/50UL);
 
 /* Bits in as_io_context.state */
 enum as_io_states {
@@ -869,7 +869,7 @@
 			if (test_bit(AS_TASK_IORUNNING, &aic->state)
 							&& in_flight == 0) {
 				thinktime = jiffies - aic->last_end_request;
-				thinktime = min(thinktime, MAX_THINKTIME-1);
+				thinktime = min(thinktime, maxthinktime-1);
 			} else
 				thinktime = 0;
 			as_update_thinktime(ad, aic, thinktime);
@@ -1951,6 +1951,7 @@
 {									\
 	return as_var_show(__VAR, (page));			\
 }
+SHOW_FUNCTION(as_maxthinktime_show, maxthinktime);
 SHOW_FUNCTION(as_readexpire_show, ad->fifo_expire[REQ_SYNC]);
 SHOW_FUNCTION(as_writeexpire_show, ad->fifo_expire[REQ_ASYNC]);
 SHOW_FUNCTION(as_anticexpire_show, ad->antic_expire);
@@ -1968,6 +1969,7 @@
 		*(__PTR) = (MAX);					\
 	return ret;							\
 }
+STORE_FUNCTION(as_maxthinktime_store, &maxthinktime, 0, LONG_MAX);
 STORE_FUNCTION(as_readexpire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
 STORE_FUNCTION(as_writeexpire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
 STORE_FUNCTION(as_anticexpire_store, &ad->antic_expire, 0, INT_MAX);
@@ -1977,6 +1979,11 @@
 			&ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
 #undef STORE_FUNCTION
 
+static struct as_fs_entry as_maxthinktime = {
+	.attr = {.name = "maxthinktime", .mode = S_IRUGO | S_IWUSR },
+	.show = as_maxthinktime_show,
+	.store = as_maxthinktime_store,
+};
 static struct as_fs_entry as_est_entry = {
 	.attr = {.name = "est_time", .mode = S_IRUGO },
 	.show = as_est_show,
@@ -2008,6 +2015,7 @@
 };
 
 static struct attribute *default_attrs[] = {
+	&as_maxthinktime.attr,
 	&as_est_entry.attr,
 	&as_readexpire_entry.attr,
 	&as_writeexpire_entry.attr,
#3 lore raw
Quoted 4 lines · FabF

I notice huge variations in first seconds of a 10 client
throughput activity as attached (5,100,300 as maxthinktime). It's just another parameter I'd like to play with ;)

Fair enough. Interesting result.

Can you describe and/or publish ffbench?

#4 lore raw
In reply to FabF
Quoted 11 lines · FabF
Quoted 5 lines · Andrew Morton

Badari, did you find any need to vary this in the AS tuning work which you were doing? (What happened to that, btw?)

If we're going to expose this tunable to users it needs to be documented in as-iosched.txt

I'm an absolute beginner Andrew ... Maybe someone more experienced could do that ?

No no, writing end user documentation is much easier than writing kernel code!

But either way you needn't worry about it because this isn't something that you would want to tweak from userspace really. If you find it useful to do some tuning then I guess it can go in -mm.

Quoted 8 lines

PS:Attached the patch v2.

Best regards,
Fabian

------------------------------------------------------------------------

Quoted 8 lines

ffbench 11 - Fabian Frederick 04/2004 - asio - maxthinktime - 10 clients
5 100 300
---------------------
5.37 21.26 16.28
11.71 23.10 20.14
15.22 21.05 20.20
13.85 18.03 17.55
11.58 16.39 16.86

...

maxthinktime of 5 is below the default antic_expire of 6. This basically turns off the thinktime heuristic, so it isn't really surprising that it runs like crap.

The *minimum* maxthinktime you should be playing with is probably about double the antic_expire value.

Raising the value will make the heuristic a bit more restrictive which could be a good idea. I think OSDL's STP has a couple of database tests where the anticipatory scheduler has a small regression (dbt2, dbt3). They might be helpful to you.