[
https://issues.apache.org/jira/browse/HBASE-17932?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15972288#comment-15972288
]
xinxin fan commented on HBASE-17932:
------------------------------------
@Allan Yang , your idea is right. The throughputs of both clients always
arise(drops) at the same time, it implies that the troughs in Figure1 really
result of GCs. And i test the case1( test big updates only) and case2(use 150
handlers for client1 and 3 handler for client2) you have mentioned, both
results exist troughs.
As for HBase 2.0, i will test the case below:
1. test big updates only and observe the active handler number
2. first run small updates , then run big updates. observe the throughput of
the small updates
we haven't develop the feature of 'manageable call queue' , so it's impossible
to do a A/B tests.
In addition, i will do a A/B test in another scenario: both client-1 and
client-2 send get requests, the client-1 search all data from lots of hfiles(
requestdistribution=uniform), the system io would be very busy. the client-2 do
not require any I/O resources(say, data is cached). The blockcache will be
bucketcache with offheap mode.
A test:
both clients send requests to a single queue, observe the throughput of both
clients
B test:
client-1 sends requests to queue1 and client-2 sends requests to queue2,
observe the throughput of both client
> add manageable call queue
> -------------------------
>
> Key: HBASE-17932
> URL: https://issues.apache.org/jira/browse/HBASE-17932
> Project: HBase
> Issue Type: Improvement
> Components: IPC/RPC
> Reporter: xinxin fan
> Fix For: 1.1.2
>
> Attachments: ManageableCallQueue.pdf
>
>
> The Manageable Call Queue aims to provide a unified way to build:
> 1. Administor can create queue with specified number handler using shell
> command
> 2. Administor can increase/decrease the handler number for the specified
> queue using shell command
> 3. Administor can assign the same type request load to one queue, for
> example, administor can assign all the large object write requests to one
> queue, while assign all the small object write requests to another queue.
> the feature allow you to change the ipc queues/handlers at runtime and is a
> improvements for running multiple workloads on a single hbase cluster in some
> cases.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
