Upgradable lock to allow simutaleous read operation while reportDiff is in
progress in processing block reports
---------------------------------------------------------------------------------------------------------------
Key: HDFS-2490
URL: https://issues.apache.org/jira/browse/HDFS-2490
Project: Hadoop HDFS
Issue Type: Sub-task
Reporter: Tomasz Nykiel
Assignee: Tomasz Nykiel
Attachments: FSNamesystemLock.java
Currently, FSNamesystem operations are protected by a single
ReentrantReadWriteLock, which allows for having multiple concurrent readers to
perform reads, and a single writer to perform writes. There are, however,
operations whose execution has primarily reading nature, but occasionally they
write.
The finest example is processing block reports - currently the entire
processing is done under writeLock(). With HDFS-395 (explicit deletion acks),
processing a block report is primarily a read operation (reportDiff()) after
which only very few blocks need to be updated. In fact, we noticed this number
to be very low, or even zero blocks.
It would be desirable to have an upgradeable read lock, which would allow for
performing other reads during the first "read" part of reportDiff() (and
possibly other operations.
We implemented such mechanism, which provides writeLock(), readLock(),
upgradeableReadLock, upgradeLock(), and downgradeLock(). I achieved this be
emloying two ReentrantReadWriteLock's - one protects writes (lock1), the other
one reads (lock2).
Hence, we have:
writeLock()
lock1.writeLock().lock()
lock2.writeLock().lock()
readLock()
lock2.readLock().lock()
upgradeableReadLock()
lock1.writeLock().lock()
upgrade()
lock2.writeLock().lock()
--------------------------
Hence a writeLock() is essentially equivalent to upgradeableLock()+upgrade()
- two writeLocks are mutually exclusive because of lock1.writeLock
- a writeLock and upgradeableLock are mutually exclusive as above
- readLock is mutually exclusive with upgradeableLock()+upgrade() OR writeLock
because of lock2.writeLock
- readLock() + writeLock() causes a deadlock, the same as currently
- writeLock() + readLock() does not cause deadlocks
--------------------------
I am conviced to the soundness of this mechanism.
The overhead comes from having two locks, and in particular, writes need to
acquire both of them.
We deployed this feature, we used the upgradeableLock() ONLY for processing
reports.
Our initial, but not exhaustive experiments have shown that it had a very
detrimental effect on the NN throughput - writes were taking up to twice as
long.
This is very unexpected, and hard to explain by only the overhead of acquiring
additional lock for writes.
I would like to ask for input, as maybe I am missing some fundamental problem
here.
I am attaching a java class which implements this locking mechanism.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
