hoping the new version coming soon

[979357361]

@daanx, hi daanx, I see that the last version is released in May, is there any news about the next release?

When I use mimalloc with pure 1GB hugepages, it comes to a 40% memory usage higher than nomal 4KB page, and with normal 4KB page, the RSS of mimalloc is still 10% higher than jemalloc, is there any clue? I have to let RocksDB work on hugepage, so mimalloc is the only choice.

I trust mi and je are both SOTA mallocs, so hoping the next version can improve this.

[daanx]

Well, a fresh version is coming soon v1.8.8 and v2.1.8 -- hopefully next week. However, with huge pages there will always be more memory usage since it is a 1 GiB huge page pinned in memory -- we cannot purge any memory inside such huge pages :-) . Also, a 10% difference with respect to some other allocator can always be the case -- we are trying but it often depends on workloads etc, and sometimes memory usage is in tension with scalable performance.

(ps. having said this, you might want to try out the latest dev3-bin branch which has a new backend allocation mechanism which under certain workloads uses much less memory. This is still in alpha stage though).

[979357361]

Well, a fresh version is coming soon v1.8.8 and v2.1.8 -- hopefully next week.

• @daanx , thanks for your reply. Glad to here that!

However, with huge pages there will always be more memory usage since it is a 1 GiB huge page pinned in memory -- we cannot purge any memory inside such huge pages :-) . -That sounds like a problem. We will repeatedly alloc and free same size memory block, if a block is freed and waiting purge (which can't be done in a hugepage), then another malloc call with same size comes, will the freed block be re-used first for allocation? If not, it sounds like we have lost the block forever, and hugepages will become unusable as time goes by? Is there any setting to improve the efficiency of hugepages?

(ps. having said this, you might want to try out the latest dev3-bin branch which has a new backend allocation mechanism which under certain workloads uses much less memory. This is still in alpha stage though). -We has tried this patch and it seems not help.