RE: Speed up transaction completion faster after many relations are accessed in a transaction

From: Alvaro Herrera [mailto:alvherre@2ndquadrant.com]
> On 2019-Sep-03, Tsunakawa, Takayuki wrote:
> > I don't think it's rejected.  It would be a pity (mottainai) to refuse
> > this, because it provides significant speedup despite its simple
> > modification.
> 
> I don't necessarily disagree with your argumentation, but Travis is
> complaining thusly:

I tried to revise David's latest patch (v8) and address Tom's comments in his last mail.  But I'm a bit at a loss.

First, to accurately count the maximum number of acquired locks in a transaction, we need to track the maximum entries in the hash table, and make it available via a new function like hash_get_max_entries().  However, to cover the shared partitioned hash table (that is not necessary for LockMethodLocalHash), we must add a spin lock in hashhdr and lock/unlock it when entering and removing entries in the hash table.  It spoils the effort to decrease contention by hashhdr->freelists[].mutex.  Do we want to track the maximum number of acquired locks in the global variable in lock.c, not in the hash table?

Second, I couldn't understand the comment about the fill factor well.  I can understand that it's not correct to compare the number of hash buckets and the number of locks.  But what can we do?


I'm sorry to repeat what I mentioned in my previous mail, but my v2 patch's approach is based on the database textbook and seems intuitive.  So I attached the rebased version.


Regards
Takayuki Tsunakawa