|Jeff Davis <pgsql(at)j-davis(dot)com>
|Memory Accounting v11
|Raw Message | Whole Thread | Download mbox | Resend email
This patch tracks memory usage (at the block level) for all memory
contexts. Individual palloc()s aren't tracked; only the new blocks
allocated to the memory context with malloc().
It also adds a new function, MemoryContextMemAllocated() which can
either retrieve the total allocated for the context, or it can recurse
and sum up the allocations for all subcontexts as well.
This is intended to be used by HashAgg in an upcoming patch that will
bound the memory and spill to disk.
Previous discussion here:
* There was a slowdown reported of around 1-3% (depending on the exact
version of the patch) on an IBM power machine when doing an index
rebuild. The results were fairly noisy for me, but it seemed to hold up.
* Adding a struct field to MemoryContextData may be bad for the CPU
caching behavior, and may be the cause of the above slowdown.
* Previous versions of the patch updated the parent contexts'
allocations as well, which avoided the need to recurse when querying the
total allocation. That seemed to penalize cases that didn't need to
track the allocation. We discussed trying to "opt-in" to this behavior,
but it seemed more awkward than helpful. Now, the patch only updates the
allocation of the current context, and querying means recursing through
the child contexts.
* There was a concern that, if MemoryContextMemAllocated needs to
recurse to the child contexts, it will be too slow for HashAgg of
array_agg, which creates a child context per group. That was solved with
My general answer to the performance concerns is that they aren't a
reason to block this patch, unless someone has a suggestion about how to
fix them. Adding one field to a struct and a few arithmetic operations
for each malloc() or realloc() seems reasonable to me.
The current state, where HashAgg just blows up the memory, is just not
reasonable, and we need to track the memory to fix that problem. Others
have also mentioned that we might want to use this mechanism to track
memory for other operators, like Sort or HashJoin, which might be
simpler and more accurate.
|Re: Memory Accounting v11
|The real reason why TAP testing isn't ready for prime time