Potential heapify performance issue in dm-vdo

Potential heapify performance issue in dm-vdo

[Kuan-Wei Chiu]

Hi Matthew,

Recently noticed that the current heapify method in min_heap.h may
degrade performance when the heap contains many compare-equal elements,
compared to the previous version.

In detail, the new heapify reduces the number of comparisons by about
50% when all elements are distinct. However, when all elements are
equal, the comparison count can degrade from O(1) to O(log n).

I don't have enough domain knowledge of dm-vdo, so I'd like to ask
whether it uses heaps with many compare-equal elements. If so, I'll
work on fixing the issue.

Regards,
Kuan-Wei

Re: Potential heapify performance issue in dm-vdo

[Matthew Sakai]

On 5/21/25 11:04 AM, Kuan-Wei Chiu wrote:
> Hi Matthew,
> 
> Recently noticed that the current heapify method in min_heap.h may
> degrade performance when the heap contains many compare-equal elements,
> compared to the previous version.
> 
> In detail, the new heapify reduces the number of comparisons by about
> 50% when all elements are distinct. However, when all elements are
> equal, the comparison count can degrade from O(1) to O(log n).
> 
> I don't have enough domain knowledge of dm-vdo, so I'd like to ask
> whether it uses heaps with many compare-equal elements. If so, I'll
> work on fixing the issue.
> 
> Regards,
> Kuan-Wei

Hi Kuan-Wei,

dm-vdo uses heapify for two different operations, but in both cases we 
define heap elements that can never be equal to each other. So I think 
this is not an issue for dm-vdo. (We have not noticed any issues with 
this in our regular testing, either.)

Matt

Re: Potential heapify performance issue in dm-vdo

[Kuan-Wei Chiu]

On Wed, May 21, 2025 at 02:57:26PM -0400, Matthew Sakai wrote:
> 
> 
> On 5/21/25 11:04 AM, Kuan-Wei Chiu wrote:
> > Hi Matthew,
> > 
> > Recently noticed that the current heapify method in min_heap.h may
> > degrade performance when the heap contains many compare-equal elements,
> > compared to the previous version.
> > 
> > In detail, the new heapify reduces the number of comparisons by about
> > 50% when all elements are distinct. However, when all elements are
> > equal, the comparison count can degrade from O(1) to O(log n).
> > 
> > I don't have enough domain knowledge of dm-vdo, so I'd like to ask
> > whether it uses heaps with many compare-equal elements. If so, I'll
> > work on fixing the issue.
> > 
> > Regards,
> > Kuan-Wei
> 
> Hi Kuan-Wei,
> 
> dm-vdo uses heapify for two different operations, but in both cases we
> define heap elements that can never be equal to each other. So I think this
> is not an issue for dm-vdo. (We have not noticed any issues with this in our
> regular testing, either.)
>
Thanks for confirming that dm-vdo won't encounter this issue.

We observed a regression in bcache caused by this behavior, and Coly
was concerned that dm-vdo might be affected in a similar way.
Fortunately, that doesn't seem to be the case.

Regards,
Kuan-Wei