[perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[Akira Yokosawa]

Hi Paul,

I find the analogy of Figure 7.11 hard to grasp.

Whether a lock is global or per-instance, the cost of locking
(saw kerf) is observed only when a CPU/thread does the locking
operation.

In this figure, does each board represent data elements, not a
CPU/thread?  If this is the case, what does the waste of "saw kerf"
mean?

What am I missing?

(I hope I am clear enough on what I don't get...)

        Thanks, Akira

Re: [perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[Paul E. McKenney]

On Sat, Jun 25, 2022 at 08:12:09AM +0900, Akira Yokosawa wrote:
> Hi Paul,
> 
> I find the analogy of Figure 7.11 hard to grasp.
> 
> Whether a lock is global or per-instance, the cost of locking
> (saw kerf) is observed only when a CPU/thread does the locking
> operation.
> 
> In this figure, does each board represent data elements, not a
> CPU/thread?  If this is the case, what does the waste of "saw kerf"
> mean?
> 
> What am I missing?
> 
> (I hope I am clear enough on what I don't get...)

It might well be that I am getting too excited about this one.  ;-)

Maybe I need to drop it.  At the very least, I need to much more clearly
explain it.

But...

Each board represents one lock.  The "saw kerf" is the time lost when
releasing that lock and someone else immediately acquiring it.

Does that help?

							Thanx, Paul

Re: [perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[Akira Yokosawa]

On Fri, 24 Jun 2022 16:33:16 -0700, Paul E. McKenney wrote:
> On Sat, Jun 25, 2022 at 08:12:09AM +0900, Akira Yokosawa wrote:
>> Hi Paul,
>>
>> I find the analogy of Figure 7.11 hard to grasp.
>>
>> Whether a lock is global or per-instance, the cost of locking
>> (saw kerf) is observed only when a CPU/thread does the locking
>> operation.
>>
>> In this figure, does each board represent data elements, not a
>> CPU/thread?  If this is the case, what does the waste of "saw kerf"
>> mean?
>>
>> What am I missing?
>>
>> (I hope I am clear enough on what I don't get...)
> 
> It might well be that I am getting too excited about this one.  ;-)
> 
> Maybe I need to drop it.  At the very least, I need to much more clearly
> explain it.
> 
> But...
> 
> Each board represents one lock.  The "saw kerf" is the time lost when
> releasing that lock and someone else immediately acquiring it.
> 
> Does that help?

Well then, why does the left side figure have ten boards?

        Thanks, Akira

> 
> 							Thanx, Paul

Re: [perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[Paul E. McKenney]

On Sat, Jun 25, 2022 at 08:57:35AM +0900, Akira Yokosawa wrote:
> On Fri, 24 Jun 2022 16:33:16 -0700, Paul E. McKenney wrote:
> > On Sat, Jun 25, 2022 at 08:12:09AM +0900, Akira Yokosawa wrote:
> >> Hi Paul,
> >>
> >> I find the analogy of Figure 7.11 hard to grasp.
> >>
> >> Whether a lock is global or per-instance, the cost of locking
> >> (saw kerf) is observed only when a CPU/thread does the locking
> >> operation.
> >>
> >> In this figure, does each board represent data elements, not a
> >> CPU/thread?  If this is the case, what does the waste of "saw kerf"
> >> mean?
> >>
> >> What am I missing?
> >>
> >> (I hope I am clear enough on what I don't get...)
> > 
> > It might well be that I am getting too excited about this one.  ;-)
> > 
> > Maybe I need to drop it.  At the very least, I need to much more clearly
> > explain it.
> > 
> > But...
> > 
> > Each board represents one lock.  The "saw kerf" is the time lost when
> > releasing that lock and someone else immediately acquiring it.
> > 
> > Does that help?
> 
> Well then, why does the left side figure have ten boards?

Ten locks.  For example, the single board might correspond to a hash
table guarded by a single global lock.  The ten boards might correspond
to a hash table with ten buckets, with per-bucket locking.

But it is sounding like this analogy might be more confusing than
enlightening.

							Thanx, Paul

Re: [perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[Akira Yokosawa]

On Fri, 24 Jun 2022 17:30:53 -0700, Paul E. McKenney wrote:
> On Sat, Jun 25, 2022 at 08:57:35AM +0900, Akira Yokosawa wrote:
>> On Fri, 24 Jun 2022 16:33:16 -0700, Paul E. McKenney wrote:
>>> On Sat, Jun 25, 2022 at 08:12:09AM +0900, Akira Yokosawa wrote:
>>>> Hi Paul,
>>>>
>>>> I find the analogy of Figure 7.11 hard to grasp.
>>>>
>>>> Whether a lock is global or per-instance, the cost of locking
>>>> (saw kerf) is observed only when a CPU/thread does the locking
>>>> operation.
>>>>
>>>> In this figure, does each board represent data elements, not a
>>>> CPU/thread?  If this is the case, what does the waste of "saw kerf"
>>>> mean?
>>>>
>>>> What am I missing?
>>>>
>>>> (I hope I am clear enough on what I don't get...)
>>>
>>> It might well be that I am getting too excited about this one.  ;-)
>>>
>>> Maybe I need to drop it.  At the very least, I need to much more clearly
>>> explain it.
>>>
>>> But...
>>>
>>> Each board represents one lock.  The "saw kerf" is the time lost when
>>> releasing that lock and someone else immediately acquiring it.
>>>
>>> Does that help?
>>
>> Well then, why does the left side figure have ten boards?
> 
> Ten locks.  For example, the single board might correspond to a hash
> table guarded by a single global lock.  The ten boards might correspond
> to a hash table with ten buckets, with per-bucket locking.

Ah, I think I understand what you mean.

On the "Global" side, there is a per-resource locking mechanism
implemented under the protection of a global lock, whereas locks
on the "Per-Instance" side are implemented independently with
each other.

Different boards can be acquired in parallel on both sides of the
figure.

Am I on the same page with you now?

> 
> But it is sounding like this analogy might be more confusing than
> enlightening.

An analogy which needs a lot of explanation might not be a good
analogy...

        Thanks, Akira

> 
> 							Thanx, Paul

Re: [perfbook] Analogy of Figure 7.11 Locking “Saw Kerf”

[Paul E. McKenney]

On Sun, Jun 26, 2022 at 01:34:53PM +0900, Akira Yokosawa wrote:
> On Fri, 24 Jun 2022 17:30:53 -0700, Paul E. McKenney wrote:
> > On Sat, Jun 25, 2022 at 08:57:35AM +0900, Akira Yokosawa wrote:
> >> On Fri, 24 Jun 2022 16:33:16 -0700, Paul E. McKenney wrote:
> >>> On Sat, Jun 25, 2022 at 08:12:09AM +0900, Akira Yokosawa wrote:
> >>>> Hi Paul,
> >>>>
> >>>> I find the analogy of Figure 7.11 hard to grasp.
> >>>>
> >>>> Whether a lock is global or per-instance, the cost of locking
> >>>> (saw kerf) is observed only when a CPU/thread does the locking
> >>>> operation.
> >>>>
> >>>> In this figure, does each board represent data elements, not a
> >>>> CPU/thread?  If this is the case, what does the waste of "saw kerf"
> >>>> mean?
> >>>>
> >>>> What am I missing?
> >>>>
> >>>> (I hope I am clear enough on what I don't get...)
> >>>
> >>> It might well be that I am getting too excited about this one.  ;-)
> >>>
> >>> Maybe I need to drop it.  At the very least, I need to much more clearly
> >>> explain it.
> >>>
> >>> But...
> >>>
> >>> Each board represents one lock.  The "saw kerf" is the time lost when
> >>> releasing that lock and someone else immediately acquiring it.
> >>>
> >>> Does that help?
> >>
> >> Well then, why does the left side figure have ten boards?
> > 
> > Ten locks.  For example, the single board might correspond to a hash
> > table guarded by a single global lock.  The ten boards might correspond
> > to a hash table with ten buckets, with per-bucket locking.
> 
> Ah, I think I understand what you mean.
> 
> On the "Global" side, there is a per-resource locking mechanism
> implemented under the protection of a global lock, whereas locks
> on the "Per-Instance" side are implemented independently with
> each other.
> 
> Different boards can be acquired in parallel on both sides of the
> figure.
> 
> Am I on the same page with you now?

Exactly!

> > But it is sounding like this analogy might be more confusing than
> > enlightening.
> 
> An analogy which needs a lot of explanation might not be a good
> analogy...

Agreed.  So either I figure out a way to ramp an explanation into
it, or I drop it.

One way to ramp an explanation into it is to discuss locking overheads
using hash tables, then go from there.  Though at that point, I am
not sure how attractive the analogy would be, useful thought it has
been to me for these many years.

Thank you for reporting this problem, and I will give it some thought.

							Thanx, Paul