ISPN-1017 - HotRod client test suite hangs eating 100% CPU #224
Conversation
| @@ -31,7 +32,7 @@ import static org.infinispan.factories.KnownComponentNames.MODULE_COMMAND_INITIA | |||
| public class ComponentRegistry extends AbstractComponentRegistry { | |||
|
|
|||
| // cached component scopes | |||
| private static final Map<Class, Scopes> componentScopeLookup = new HashMap<Class, Scopes>(); | |||
| private static final Map<Class, Scopes> componentScopeLookup = Collections.synchronizedMap(new HashMap<Class, Scopes>()); | |||
There was a problem hiding this comment.
Why not use ConcurrentHashMap for better concurrency ?
There was a problem hiding this comment.
I always had a bias against ConcurrentHashMap, it seemed too heavyweight
with its support for concurrent writers. Of course, it turns up my solution
was forcing the serialization of annotation parsing as well, so it wasn't
better at all.
My version would have made some sense if there were a lot of reads after the
map was built. But my assumptions were wrong, I ran the profiler and there
are practically no calls to isGlobal after startup - which means my
optimization is actually a deoptimization.
I didn't see duplicate annotation parsing to a perf hotspot either, so I'll
give up synchronization for the annotation parsing and I'll give
ConcurrentHashMap a try.
On Apr 1, 2011 1:21 PM, "mmarkus" <
reply@reply.github.com>
wrote:
@@ -31,7 +32,7 @@ import static
org.infinispan.factories.KnownComponentNames.MODULE_COMMAND_INITIA
public class ComponentRegistry extends AbstractComponentRegistry {// cached component scopes
- private static final Map<Class, Scopes> componentScopeLookup = new
HashMap<Class, Scopes>();- private static final Map<Class, Scopes> componentScopeLookup =
Collections.synchronizedMap(new HashMap<Class, Scopes>());Why not use ConcurrentHashMap for better concurrency ?
Reply to this email directly or view it on GitHub:
https://github.com/infinispan/infinispan/pull/224/files#r14555
|
hi, I'm having the same issues, just starting Infinispan. I think this is quite urgent & critical. Changing this to a CHM fixed it. |
Concurrent put() calls in the component scopes HashMap were causing get() to enter an infinite loop. I changed the map to a ConcurrentHashMap to ensure atomic updates. We can still waste work and parse the annotations of a class twice, but it's not that expensive.
|
thanks! |
Updated to only use SynchronizedMap, it's not a perf hotspot.
Please also merge t_ispn1017_4.2.x in the 4.2.x branch.