marco polo created ACCUMULO-3508:
------------------------------------
Summary: Zoocache utilizes course grained locking causing
unnecessary synchronization in many tserver ops
Key: ACCUMULO-3508
URL: https://issues.apache.org/jira/browse/ACCUMULO-3508
Project: Accumulo
Issue Type: Sub-task
Components: fate
Affects Versions: 1.6.0
Reporter: marco polo
Priority: Minor
Fix For: 1.6.3
Through jstacking I have observed busy tablet server threads waiting for the
Zoocache to be updated or even read. Profiling via yourkit indicates potential
deadlocks when threads wait for a long period of time ( in my case my waits
were 10+s ) when many scans were running against my instance.
My patch, which is forthcoming, removes synchronization and utilizes a
reentrant read/write lock (RERWL) to synchronize the maps. My initial design
utilized concurrent hash maps, but the design of existence becomes simpler and
changes fewer by using the RERWL.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
