improved perf_test to measure steady state behavior

[nbds] / todo

diff --git a/todo b/todo
index ab8c0db4bca300fbedcd8a5e6e4a9af0a1c0dc88..ad68a8e294aecdc248d75c90f2c65c51f3986abe 100644 (file)
--- a/todo
+++ b/todo
@@ -1,5 +1,32 @@
-- make rcu wait when its buffer gets full, instead of crashing
-- fix makefile to compute dependency info as a side-effect of compilation (-MF)
+memory reclamation
+------------------
+- allow threads to dynamically enter and exit rcu's token passing ring
+- augment rcu with heartbeat manager to kill and recover from stalled threads
+- make rcu try yielding when its buffer gets full
+- use alternate memory reclamation schemes: hazard pointers and/or reference counting
+
+quality
+-------
+- verify the memory management of keys in list, skiplist, and hashtable
+- transaction tests
+- port perf tests from lib-high-scale
+- characterize the performance of hashtable vs. skiplist vs. list
+- validate function arguments in interface functions
+- document usage
+- document algorithms
+
+optimization
+------------
 - investigate 16 byte CAS; ht can store GUIDs inline instead of pointers to actual keys 
-- test ht
-- optimize tracing code, still too much overhead
+- txn write after write can just update the old update record instead of pushing a new one
+- use a shared scan for write-set validation in txn, similar to ht copy logic
+- experiment with the performance impact of not passing the hash between functions in ht
+- experiment with embedding the nstring keys in the list/skiplist nodes
+
+features
+--------
+- allow values of 0 to be inserted into maps (change DOES_NOT_EXIST to something other than 0)
+- read-committed type transactions
+- recycle free regions across size-classes and between threads
+
+