From 987116e8cea77ef67ee49495198e2150d14a933a Mon Sep 17 00:00:00 2001 From: cc32d9 Date: Fri, 26 Aug 2022 14:31:01 +0200 Subject: [PATCH] Update mapped_usability.md --- chainbase/mapped_usability.md | 3 +++ 1 file changed, 3 insertions(+) diff --git a/chainbase/mapped_usability.md b/chainbase/mapped_usability.md index 98d9235..2e4e794 100644 --- a/chainbase/mapped_usability.md +++ b/chainbase/mapped_usability.md @@ -24,3 +24,6 @@ Obviously relying on undocumented behavior is yucky. But realistically the risk This could be mitigated by including a fixed version of boost within chainbase. Something we've noodled on a little before and may be required for performance improvement in other proposal. +### use process RAM for uncommitted changes + +Currently chainbase uses the mmap memory for everything, including uncommitted changes that are coming with speculative transactions. But the uncommitted changes don't have to be in the state memory, and it makes sense to store them in normal malloc memory. This would reduce the I/O load on the state storage significantly, so that we would probably not need tmpfs any longer.