[
https://issues.apache.org/jira/browse/JEXL-414?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17791019#comment-17791019
]
Holger Sunke commented on JEXL-414:
-----------------------------------
Just wanted to add that in our case the SoftCache was configured too small
(capacity 5000 where 10000 was needed) and the write-lock actually figured out
to be the bottleneck during throughput peaks. I pulled threaddumps when things
got wild (giant response times) and found hundreds of threads waiting for a
lock on the cache which I didn't even know is existing up to that moment.
Increasing the cache size boosted overall response times visibly in all
monitoring graphs. This is likely caused due to less expression parsing but
maybe also less (exclusive write-)locking.
At least in situations where the app is close to congestion due to a peak I'm
sure the SyncronizedMap would be a problem in our case. Thus the solution of
now being able to choose and experiment on different Map implementations is
great :)
> SoftCache may suffer from race conditions
> -----------------------------------------
>
> Key: JEXL-414
> URL: https://issues.apache.org/jira/browse/JEXL-414
> Project: Commons JEXL
> Issue Type: Bug
> Affects Versions: 3.3
> Reporter: Holger Sunke
> Assignee: Henri Biestro
> Priority: Major
> Fix For: 3.3.1
>
>
> I found the SoftCache used from within the JEXL class Engine to be very
> relevant for overall performance (depending on the application) on the one
> hand, but on the other hand it may suffer from race conditions.
> While solid efford was taken to protect it from race conditions by
> surrounding access using a ReadWriteLock, parallel read access actually do
> reach out on a LinkedHashMap having set accessOrder=true. This means that on
> every potentially parallel, non serialized invocation of LinkedHashMap#get
> the order of elements is modified to bring the last accessed element down to
> the tail of the internal linked list structure and modCount is incremented.
>
> In our application rendering web templates, we observe multiple of 10000
> access on the SoftCache per page impression and multiple of 10 page
> impressions per second per JVM instance.
> Not sure whether this is result of race condition claimed above, in heapdumps
> of that application I observed the LinkedHashMap used within SoftCache having
> a size of ~9500 elements while the SoftCache#size is set to limit the cache
> to 5000 elements. Additionaly the LinkedHashMap#modCount shows up with
> arbitrary giant positive or negative numbers, telling me it has overflown
> already after some hours of application uptime.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
