[
https://issues.apache.org/jira/browse/HDFS-12838?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16338178#comment-16338178
]
Anu Engineer commented on HDFS-12838:
-------------------------------------
[~msingh] Thanks for investigating the cache. This patch looks quite good. I
have only one comment.
* KeyManagerImpl.java:allocateBlocks*
{code}
while (requestedSize > 0) {
long allocateSize = Math.min(scmBlockSize, requestedSize);
sizes.add(allocateSize);
requestedSize -= allocateSize;
}
{code}
I am presuming that we are trying to do here it to convert
the requested size to blocks. For example, if the block size is 128M and user
requested 500M of space, we will convert that into 500MB / 128MB = 4 blocks,
yielding an allocation in SCM of 4 blocks.
If my assumption is true, then would it make sense to
rewrite the following API
{{scmBlockClient.allocateBlock(sizes, type, factor, ksmId);}}
as
{{scmBlockClient.allocateBlock(blockSize, blockCount, type, factor, ksmId);}}
and now the original while loop can be replaced as
{code}
int blockCount = (requestedSize % scmBlockSize) + 1;
scmBlockClient.allocateBlock(scmBlockSize, blockCount, type, factor, ksmId);
{code}
This does mean that client will have to overwrite the actual block length after
the update is done.
This means that the signature in {{ScmBlockLocationProtocol.java}} would have
to change too
and ScmBlockLocationProcotol.proto -- would change to
{{uint64 size = 1;}} and {{uint32 count = 2;}}
* nit: ScmBlockLocationProtocolClientSideTranslatorPB, seems like we have an
unused import.
> Ozone: Optimize number of allocated block rpc by aggregating multiple block
> allocation requests
> -----------------------------------------------------------------------------------------------
>
> Key: HDFS-12838
> URL: https://issues.apache.org/jira/browse/HDFS-12838
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: ozone
> Affects Versions: HDFS-7240
> Reporter: Mukul Kumar Singh
> Assignee: Mukul Kumar Singh
> Priority: Major
> Fix For: HDFS-7240
>
> Attachments: HDFS-12838-HDFS-7240.001.patch,
> HDFS-12838-HDFS-7240.002.patch, HDFS-12838-HDFS-7240.003.patch,
> HDFS-12838-HDFS-7240.004.patch, HDFS-12838-HDFS-7240.005.patch,
> HDFS-12838-HDFS-7240.006.patch
>
>
> Currently KeySpaceManager allocates multiple blocks by sending multiple block
> allocation requests over the RPC. This can be optimized to aggregate multiple
> block allocation request over one rpc.
> {code}
> while (requestedSize > 0) {
> long allocateSize = Math.min(scmBlockSize, requestedSize);
> AllocatedBlock allocatedBlock =
> scmBlockClient.allocateBlock(allocateSize, type, factor);
> KsmKeyLocationInfo subKeyInfo = new KsmKeyLocationInfo.Builder()
> .setContainerName(allocatedBlock.getPipeline().getContainerName())
> .setBlockID(allocatedBlock.getKey())
> .setShouldCreateContainer(allocatedBlock.getCreateContainer())
> .setIndex(idx++)
> .setLength(allocateSize)
> .setOffset(0)
> .build();
> locations.add(subKeyInfo);
> requestedSize -= allocateSize;
> }
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
