Hi, all. In the last week, I replaced linked list with hash table in 
 I only replace inConflicts and outConflicts. The other linked lists, such as 
possibleUnsafeConflicts, I will modify them after other things work well.

There are still some bugs: the abort rate is much higher than before (from 
11.3% to 72%). I'll figure out what's wrong in the next few days.

My design is as follow:

For hash table, key is the pointer of SerializableXact; Value is the 
RWConflictData object. 
Hashcode is generated based on the SerializableXact pointer. 
So, given a SerializableXact, we can quickly find if it is conflict with 
another SerializableXact.

Every SerializableXact has two tables: inConflictsTab and outConflictsTab. 
They are allocated and initialized when creating PredXactList (in the function 
When a SerializableXact object is released, it will release all its RWConflict 
objects, the hash tables are empty again. 
Thus They can be reused directly after releasing. 

NOTE: I stored RWConflictData in hash tables, instead of RWConflict object 
allocated by RWConflictPool. 
 After I remove other linked lists, the RWConflictPool can be omitted.

My code is on the :

There are 3 main modifications in summary:
1) Initializing hash table. Related functions: InitPredicateLocks

2) Setting, releasing and checking RWConflicts. Related functions: 
RWConflictExists, SetRWConflict, ReleaseRWConflict

3) There are some functions scanning all RWConflict in a SerializableXact.
Related functions: ReleasePredicateLocks, ReleaseOneSerializableXact, 
 CheckForSerializableConflictIn, OnConflict_CheckForSerializationFailure, 

Mengxing Liu

Reply via email to