[
https://issues.apache.org/jira/browse/HADOOP-15016?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16310503#comment-16310503
]
Xiaoyu Yao commented on HADOOP-15016:
-------------------------------------
[~ywskycn], thanks for the updated design doc. Some initial feedback:
1. This can be a useful feature for multi-tenancy Hadoop cluster. The cost
estimates for different RPC calls can be difficult. Instead of hardcode fixed
value per RPC, I would suggest making it a pluggable interface so that we can
customize it for different deployments.
2. The reserved share of call queue looks good. It is similar what we proposed
in HADOOP-13128. What do we plan to handle the case when the reserved queue is
full? blocking or backoff?
3. The feature might need many manual configurations and tune to work for
specific deployment and workloads. Do you want to add a section to discuss
configurations, CLI tools, etc. to make this easier to use?
4. It would be great if you could share some of the results achieved with the
POC patch (e.g., RPC/second, average locking, process and queue time with/wo
the patch).
> Cost-Based RPC FairCallQueue with Reservation support
> -----------------------------------------------------
>
> Key: HADOOP-15016
> URL: https://issues.apache.org/jira/browse/HADOOP-15016
> Project: Hadoop Common
> Issue Type: Improvement
> Reporter: Wei Yan
> Assignee: Wei Yan
> Attachments: Adding reservation support to NameNode RPC resource.pdf,
> Adding reservation support to NameNode RPC resource_v2.pdf,
> HADOOP-15016_poc.patch
>
>
> FairCallQueue is introduced to provide RPC resource fairness among different
> users. In current implementation, each user is weighted equally, and the
> processing priority for different RPC calls are based on how many requests
> that user sent before. This works well when the cluster is shared among
> several end-users.
> However, this has some limitations when a cluster is shared among both
> end-users and some service jobs, like some ETL jobs which run under a service
> account and need to issue lots of RPC calls. When NameNode becomes quite
> busy, this set of jobs can be easily backoffed and low-prioritied. We cannot
> simply treat this type jobs as "bad" user who randomly issues too many calls,
> as their calls are normal calls. Also, it is unfair to weight a end-user and
> a heavy service user equally when allocating RPC resources.
> One idea here is to introduce reservation support to RPC resources. That is,
> for some services, we reserve some RPC resources for their calls. This idea
> is very similar to how YARN manages CPU/memory resources among different
> resource queues. A little more details here: Along with existing
> FairCallQueue setup (like using 4 queues with different priorities), we would
> add some additional special queues, one for each special service user. For
> each special service user, we provide a guarantee RPC share (like 10% which
> can be aligned with its YARN resource share), and this percentage can be
> converted to a weight used in WeightedRoundRobinMultiplexer. A quick example,
> we have 4 default queues with default weights (8, 4, 2, 1), and two special
> service users (user1 with 10% share, and user2 with 15% share). So finally
> we'll have 6 queues, 4 default queues (with weights 8, 4, 2, 1) and 2 special
> queues (user1Queue weighted 15*10%/75%=2, and user2Queue weighted
> 15*15%/75%=3).
> For new coming RPC calls from special service users, they will be put
> directly to the corresponding reserved queue; for other calls, just follow
> current implementation.
> By default, there is no special user and all RPC requests follow existing
> FairCallQueue implementation.
> Would like to hear more comments on this approach; also want to know any
> other better solutions? Will put a detailed design once get some early
> comments.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
