[
https://issues.apache.org/jira/browse/HDFS-8611?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15124494#comment-15124494
]
Xiaobing Zhou commented on HDFS-8611:
-------------------------------------
Thanks [~walter.k.su] for the proposal, however Circular FIFO queue is fixed
size. This could be a problem since both PriorityQueue as member of
LightWeightCache and LightWeightGSet as base are of no upper bound regarding
number of elements they stored.
> Improve the performance of retry cache eviction
> -----------------------------------------------
>
> Key: HDFS-8611
> URL: https://issues.apache.org/jira/browse/HDFS-8611
> Project: Hadoop HDFS
> Issue Type: Improvement
> Reporter: Kihwal Lee
> Assignee: Xiaobing Zhou
> Priority: Critical
>
> As discussed in HDFS-7609, removing expired entry from retry cache can be
> costly. Following is the comment left by [~szetszwo] in HDFS-7609.
> {quote}
> PriorityQueue#remove is O\(n), so that definitely could be problematic. It's
> odd that there would be so many collisions that this would become noticeable
> though. Are any of you running a significant number of legacy applications
> linked to the RPC code before introduction of the retry cache support? If
> that were the case, then perhaps a huge number of calls are not supplying a
> call ID, and then the NN is getting a default call ID value from protobuf
> decoding, thus causing a lot of collisions.
> {quote}
> The priority queue can be improved using a balanced tree as stated in the
> java comment in LightWeightCache. We should do it if it could fix the
> problem.
> {code}
> //LightWeightCache.java
> /*
> * The memory footprint for java.util.PriorityQueue is low but the
> * remove(Object) method runs in linear time. We may improve it by using a
> * balanced tree. However, we do not yet have a low memory footprint
> balanced
> * tree implementation.
> */
> private final PriorityQueue<Entry> queue;
> {code}
> BTW, the priority queue is used to evict entries according the expiration
> time. All the entries (with any key, i.e. any caller ID) are stored in it.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
