Yi Liu created HDFS-6782:
----------------------------
Summary: Improve FS editlog logSync
Key: HDFS-6782
URL: https://issues.apache.org/jira/browse/HDFS-6782
Project: Hadoop HDFS
Issue Type: Improvement
Components: namenode
Affects Versions: 3.0.0
Reporter: Yi Liu
Assignee: Yi Liu
In NN, it uses a double buffer (bufCurrent, bufReady) for log sync, bufCurrent
it to buffer new coming edit ops and bufReady is for flushing. This's
efficient. When flush is ongoing, and bufCurrent is full, NN goes to force log
sync, and all new Ops are blocked (since force log sync is protected by
FSNameSystem write lock). After the flush finished, the new Ops are still
blocked, but actually at this time, bufCurrent is free and Ops can go ahead and
write to the buffer. The following diagram shows the detail. This JIRA is for
this improvement. Thanks [~umamaheswararao] for confirming this issue.
{code}
edit1(txid1) ------ write to bufCurrent -------- logSync --------- (swap
buffer)flushing -------
edit2(txid2) ------ write to bufCurrent -------- logSync --------- waiting
-------
edit3(txid3) ------ write to bufCurrent -------- logSync --------- waiting
-------
edit4(txid4) ------ write to bufCurrent -------- logSync --------- waiting
-------
edit5(txid5) ------ write to bufCurrent --full-- force sync --------- waiting
-------
edit6(txid6) ------ blocked
...
editn(txidn) ------ blocked
{code}
After the flush, it becomes
{code}
edit1(txid1) ------ write to bufCurrent -------- logSync --------- finished
--------
edit2(txid2) ------ write to bufCurrent -------- logSync --------- flushing
-------
edit3(txid3) ------ write to bufCurrent -------- logSync --------- waiting
-------
edit4(txid4) ------ write to bufCurrent -------- logSync --------- waiting
-------
edit5(txid5) ------ write to bufCurrent --full-- force sync --------- waiting
-------
edit6(txid6) ------ blocked
...
editn(txidn) ------ blocked
{code}
After edit1 finished, bufCurrent is free, and the thread which flushes txid2
will also flushes txid3-5, so we should return from the force sync of edit5 and
FSNamesystem write lock will be freed (Don't worry that edit5 Op will return,
since there will be a normal logSync after the force logSync and there will
wait for sync finished). This is the idea of this JIRA.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
