Skip to contentSkip to conversation

Question on updated Figure 5.1 "Atomic Increment Scalability on Kaby Lake"

PATCH In review Akira Yokosawa 2 participants 4 messages Sep 2, 2018 – Sep 12, 2018 1 message not in the archive

lore
Share & tools
lore
#0 lore raw

Hello Palik,

I'm looking at Figure 5.1 of perfbook which was updated by commit 3b62f67a76e1 ("Regenerating the atomic counter graph on a more modern CPU"), and wondering the time to increment at x = 1 looks too small.

As is described in the Answer to Quick Quiz 5.8, atomic increment should take a bit longer than the ideal non-atomic increment even when the number of updaters is 1.

Or on Kaby Lake, some special optimization results in almost no cost in this case? If that is the case, the Quick Quiz needs update...

Also, the horizontal dashed line should be closer to the x-axis, I suppose. The time of the ideal non-atomic increment is embedded on line 44 of CodeSamples/count/plots.sh as "8.81772".
You might need to use a proper value for Kaby Lake (should be near zero) instead.

If I understand the circumstances right, raw data for the previous version of the graph reside in CodeSamples/count/data/count_atomic:u.2009.05.03a.dat.

Can you share the corresponding raw data on Kaby Lake?

Also, in the above mentioned commit, atomic125.{eps|png} were overwritten accidentally. They were graphs of 125-thread case, and should not have been updated IMO. They are not used in perfbook at the moment, but I'd like to have the original ones in the repository.

If you could share the raw data for Kaby Lake, I'd be happy to fix these issues for you.

I'm not blaming any of you. I could blame the ad-hoc design of plots.sh, but it was written by Paul for his own use. Such scripts tend to be ad-hoc. ;-) I might be able to improve the script.

Thanks, Akira

A message that is not in the archive (sent off-list or not yet received): CAFMy-Rk76bc+oOXq7ijGT4wzYkFbs1F7bkLWQ-VwE_O2Q-Z2VQ@mail.gmail.com

#2 lore raw

(Added CC to Paul and perfbook list.)

Quoted 4 lines · Imre Palik

Hi Akira,

sorry for the confusion. I should have taken better care. I attached the row data from a measurement.

Thank you!

I just skimmed the output of count_nonatomic.

---
count_nonatomic 1 uperf 2
n_reads: 0 n_updates: 778797305000 nreaders: 0 nupdaters: 1 duration: 240 ns/read: -nan ns/update: 0.000308167
count_nonatomic 1 uperf 2
n_reads: 0 n_updates: 779548753000 nreaders: 0 nupdaters: 1 duration: 240 ns/read: -nan ns/update: 0.00030787
count_nonatomic 1 uperf 2
n_reads: 0 n_updates: 801753128000 nreaders: 0 nupdaters: 1 duration: 240 ns/read: -nan ns/update: 0.000299344
count_nonatomic 1 uperf 2
n_reads: 0 n_updates: 783816136000 nreaders: 0 nupdaters: 1 duration: 240 ns/read: -nan ns/update: 0.000306194
count_nonatomic 1 uperf 2
n_reads: 0 n_updates: 786531305000 nreaders: 0 nupdaters: 1 duration: 240 ns/read: -nan ns/update: 0.000305137
count_nonatomic 2 uperf 2
!!! Count mismatch: 1566325493000 counted vs. 776288032000 final value
n_reads: 0 n_updates: 1566325493000 nreaders: 0 nupdaters: 2 duration: 240 ns/read: -nan ns/update: 0.00030645
count_nonatomic 2 uperf 2
!!! Count mismatch: 1637607003000 counted vs. 825312711000 final value
n_reads: 0 n_updates: 1637607003000 nreaders: 0 nupdaters: 2 duration: 240 ns/read: -nan ns/update: 0.000293111
count_nonatomic 2 uperf 2
!!! Count mismatch: 1525890163000 counted vs. 768873049000 final value
n_reads: 0 n_updates: 1525890163000 nreaders: 0 nupdaters: 2 duration: 240 ns/read: -nan ns/update: 0.00031457
count_nonatomic 2 uperf 2
!!! Count mismatch: 1487184107000 counted vs. 717784175000 final value
n_reads: 0 n_updates: 1487184107000 nreaders: 0 nupdaters: 2 duration: 240 ns/read: -nan ns/update: 0.000322758
count_nonatomic 2 uperf 2
!!! Count mismatch: 1512515450000 counted vs. 751090603000 final value
n_reads: 0 n_updates: 1512515450000 nreaders: 0 nupdaters: 2 duration: 240 ns/read: -nan ns/update: 0.000317352
--

These results look too fast. I checked CodeSamples/count/count_nonatoimc.c and found that it has no READ_ONCE()/WRITE_ONCE() to access count.

I'm submitting a patch to fix it soon.
Can you share the raw data after applying it?

Quoted 4 lines

Thanks,

Imre

And apologies for my confusion of your first name.

Thanks, Akira

Quoted 41 lines · Akira Yokosawa
Quoted 40 lines

Hello Palik,

I'm looking at Figure 5.1 of perfbook which was updated by commit 3b62f67a76e1 ("Regenerating the atomic counter graph on a more modern CPU"), and wondering the time to increment at x = 1 looks too small.

As is described in the Answer to Quick Quiz 5.8, atomic increment should take a bit longer than the ideal non-atomic increment even when the number of updaters is 1.

Or on Kaby Lake, some special optimization results in almost no cost in this case? If that is the case, the Quick Quiz needs update...

Also, the horizontal dashed line should be closer to the x-axis, I suppose. The time of the ideal non-atomic increment is embedded on line 44 of CodeSamples/count/plots.sh as "8.81772".
You might need to use a proper value for Kaby Lake (should be near zero) instead.

If I understand the circumstances right, raw data for the previous version of the graph reside in CodeSamples/count/data/count_atomic:u.2009.05.03a.dat.

Can you share the corresponding raw data on Kaby Lake?

Also, in the above mentioned commit, atomic125.{eps|png} were overwritten accidentally. They were graphs of 125-thread case, and should not have been updated IMO. They are not used in perfbook at the moment, but I'd like to have the original ones in the repository.

If you could share the raw data for Kaby Lake, I'd be happy to fix these issues for you.

I'm not blaming any of you. I could blame the ad-hoc design of plots.sh, but it was written by Paul for his own use. Such scripts tend to be ad-hoc. ;-) I might be able to improve the script.

Thanks, Akira

#3 lore raw

PATCH 1/1 CodeSamples/count: Use READ_ONCE/WRITE_ONCE in count_nonatomic.c

From 13fa1b9b162483df5d88b7aefe408a4e17e83338 Mon Sep 17 00:00:00 2001
From: Akira Yokosawa <[email protected]>
Date: Wed, 12 Sep 2018 19:27:41 +0900
Subject: [PATCH] CodeSamples/count: Use READ_ONCE/WRITE_ONCE in count_nonatomic.c

Without them, inlined accesses can be optimized away by "-O3".

---
 CodeSamples/count/count_nonatomic.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
CodeSamples/count/count_nonatomic.c +2 -2
index 90979c5..50df475 100644
@@ -25,12 +25,12 @@ unsigned long counter = 0;
 
 __inline__ void inc_count(void)
 {
-	counter++;
+	WRITE_ONCE(counter, READ_ONCE(counter) + 1);
 }
 
 __inline__ unsigned long read_count(void)
 {
-	return counter;
+	return READ_ONCE(counter);
 }
 
 __inline__ void count_init(void)
Signature · 1 line
2.7.4
#4 lore raw

Re: [PATCH] CodeSamples/count: Use READ_ONCE/WRITE_ONCE in count_nonatomic.c

Quoted 6 lines · Akira Yokosawa
Quoted 1 line

From 13fa1b9b162483df5d88b7aefe408a4e17e83338 Mon Sep 17 00:00:00 2001

From: Akira Yokosawa <[email protected]>
Date: Wed, 12 Sep 2018 19:27:41 +0900
Subject: [PATCH] CodeSamples/count: Use READ_ONCE/WRITE_ONCE in count_nonatomic.c

Without them, inlined accesses can be optimized away by "-O3".

That code was written in a far more innocent time when compilers were far less aggressive, wasn't it?

Great catch, both on the data and on the code, thank you! Applied and pushed.

Thanx, Paul

Quoted 28 lines
---
 CodeSamples/count/count_nonatomic.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
CodeSamples/count/count_nonatomic.c +2 -2
index 90979c5..50df475 100644
@@ -25,12 +25,12 @@ unsigned long counter = 0;
 
 __inline__ void inc_count(void)
 {
-	counter++;
+	WRITE_ONCE(counter, READ_ONCE(counter) + 1);
 }
 
 __inline__ unsigned long read_count(void)
 {
-	return counter;
+	return READ_ONCE(counter);
 }
 
 __inline__ void count_init(void)
Signature · 1 line
2.7.4