hadoop-common-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sandy Ryza <sandy.r...@cloudera.com>
Subject Re: Allocating Containers on a particular Node in Yarn
Date Thu, 14 Nov 2013 20:40:34 GMT
Great to hear.  Other answers inline


On Thu, Nov 14, 2013 at 12:05 PM, Gaurav Gupta <gaurav@datatorrent.com>wrote:

> Sandy,
>
>
>
> The last trick worked but now I have couple of more questions
>
>
>
> 1.      If I don’t request for rack and relax locality is false with
> scheduler delay on , I see that when I pass a wrong host I don’t get any
> container back. Why so?
>
When locality relaxation is turned on, we will only ever place the
container on the host that you requested.  This means that if no such host
exists, we will never place the container.  This behavior could maybe be
improved, but there are some tricky issues to deal with there about nodes
going down and such.

> 2.      I also noticed that with scheduler delay on and relax locality
> set to false, If I request for rack along with a host I get one container
> on expected node but not other containers.
>
> 3.      With scheduler delay on and relax locality set to true without
> requesting rack, I don’t get the containers on the required host
>
What scheduler are you using and what properties are you using to turn the
scheduler delay on?

>
>
> Thanks
>
> -Gaurav
>
>
>
> *From:* Sandy Ryza [mailto:sandy.ryza@cloudera.com]
> *Sent:* Thursday, November 14, 2013 11:41 AM
> *To:* user@hadoop.apache.org
>
> *Subject:* Re: Allocating Containers on a particular Node in Yarn
>
>
>
> Requesting the rack is not necessary, and is leading to the behavior that
> you're seeing.
>
>
>
> The documentation states:
>
>    * <li>If locality relaxation is disabled, then only within the same
> request,
>
>    * a node and its rack may be specified together. This allows for a
> specific
>
>    * rack with a preference for a specific node within that rack.</li>
>
>
>
>
>
> On Wed, Nov 13, 2013 at 10:10 PM, gaurav <gaurav@datatorrent.com> wrote:
>
> Here is the snippet of code that I am using to allocate containers
>
>     AMRMClient<ContainerRequest> amRmClient =
> AMRMClient.createAMRMClient();;
>     String host = "h1";
>     Resource capability = Records.newRecord(Resource.class);
>     capability.setMemory(memory);
>     nodes = new String[] {host};
>     // in order to request a host, we also have to request the rack
>     racks = new String[] {"/default-rack"};
>      List<ContainerRequest> containerRequests = new
> ArrayList<ContainerRequest>();
>     List<ContainerId> releasedContainers = new ArrayList<ContainerId>();
>     containerRequests.add(new ContainerRequest(capability, nodes, racks,
> Priority.newInstance(priority),false));
>     if (containerRequests.size() > 0) {
>       LOG.info("Asking RM for containers: " + containerRequests);
>       for (ContainerRequest cr : containerRequests) {
>         LOG.info("Requested container: {}", cr.toString());
>         amRmClient.addContainerRequest(cr);
>       }
>     }
>
>     for (ContainerId containerId : releasedContainers) {
>       LOG.info("Released container, id={}", containerId.getId());
>       amRmClient.releaseAssignedContainer(containerId);
>     }
>     return amRmClient.allocate(0);
>
>
>
> -Gaurav
>
>
>
> On 11/13/2013 07:36 PM, Sandy Ryza wrote:
>
> In that case, the AMRMClient code looks correct to me.  Can you share the
> code you've written against it that's not receiving the correct containers?
>
>
>
> On Wed, Nov 13, 2013 at 5:30 PM, gaurav <gaurav@datatorrent.com> wrote:
>
> Hi,
>
> I looked in the trunk and line numbers are 361 and 366.
>
> Thanks
>
> -Gaurav
>
> On 11/13/2013 5:04 PM, gaurav wrote:
>
> I have hadoop-2.2.0
>
> Thanks
>
> -Gaurav
>
> On 11/13/2013 4:59 PM, Sandy Ryza wrote:
>
> What version are you using?  Setting the relax locality to true for nodes
> is always correct.  For racks, this is not necessarily the case.  When I
> look at trunk, it is set to true always on line 361 (which is correct), but
> on on line 374.
>
>
>
> -Sandy
>
>
>
> On Wed, Nov 13, 2013 at 4:47 PM, gaurav <gaurav@datatorrent.com> wrote:
>
> Hi Sandy,
>
> No it is not working for me.  As mentioned earlier, AMRMClient is not
> respecting the locality for node and rack in *Line 361 and 374 *and is
> set to true always.
> I am requesting for one container with specified node and rack and relaxed
> locality false.
>
> Thanks
>
> -Gaurav
>
> On 11/13/2013 4:24 PM, Sandy Ryza wrote:
>
> [moving to user list]
>
>
>
> Right.  relaxLocality needs to be set on the next level up.  It determines
> whether locality can be relaxed to that level.  Confusing, I know.  If you
> are using AMRMClient, you should be able to accomplish what you're looking
> for by creating an AMRMClient.ContainerRequest that specifies a node and
> sets relax locality to false.  Is that not working for you?
>
>
>
> -Sandy
>
>
>
> On Wed, Nov 13, 2013 at 4:19 PM, gaurav <gaurav@datatorrent.com> wrote:
>
> Hi Sandy,
>
> If I set relaxlocality to true with host name, I don't get the container
> on specified host even though node has the resources.
> I am using AMRMClient, only thing is I made following changes to get the
> containers on the specified node.
> /
> //Line: 361 //
> //Original//
> // addResourceRequest(req.getPriority(), node, req.getCapability(), req,
> true);//
> //Modifiled//
> // addResourceRequest(req.getPriority(), node, req.getCapability(), req,
> req.getRelaxLocality());//
> //
> //Line 374//
> //Original//
> // addResourceRequest(req.getPriority(), rack, req.getCapability(), req,
> true);//
> //Modifiled//
> // addResourceRequest(req.getPriority(), rack, req.getCapability(), req,
> req.getRelaxLocality());//
> ///
>
>
> Thanks
> -Gaurav
>
>
>
> On 11/13/2013 4:02 PM, Sandy Ryza wrote:
>
> Yeah, specifying a host name with relaxLocality is meaningful.  Schedulers
> use delay scheduling (
> http://www.cs.berkeley.edu/~matei/talks/2010/eurosys_delaysched.pdf) to
> achieve locality when relaxLocality is on.  But it is turned off by
> default.  The individual scheduler docs have specifics on how to configure
> it.
>
> Guarav,
> Using ResourceRequests directly is not straightforward and error prone.  Is
> there a reason that AMRMClient is unsuitable for your needs?
>
> -Sandy
>
>
>
> On Wed, Nov 13, 2013 at 3:55 PM, Thomas Weise <thomas.weise@gmail.com
> >wrote:
>
> Is it possible to specify a particular node and have RM fallback to an
> different node only after making an attempt to allocate for the requested
> node? In other words, is the combination of specific host name and
> relaxLocality=TRUE meaningful at all?
>
> Thanks.
>
>
> On Wed, Nov 13, 2013 at 3:23 PM, Alejandro Abdelnur <tucu@cloudera.com
>
> wrote:
> Gaurav,
>
> Setting relaxLocality to FALSE should do it.
>
> thanks.
>
>
> On Wed, Nov 13, 2013 at 2:58 PM, gaurav <gaurav@datatorrent.com> wrote:
>
>   Hi,
>   I am trying to allocate containers on a particular node in Yarn but
>
> Yarn
>
> is returning me containers on different node although the requested
>
> node
>
> has resources available.
>
> I checked into the allocate(AllocateRequest request) function of
> ApplicationMasterService and my request is as follows
>
> *request: ask { priority { priority: 1 } resource_name: "h2"
>
> capability {
>
> memory: 1000 } num_containers: 2 } ask { priority { priority: 1 }
> resource_name: "/default-rack" capability { memory: 1000 }
>
> num_containers:
>
> 2 } ask { priority { priority: 1 } resource_name: "*" capability {
>
> memory:
>
> 1000 } num_containers: 2 } response_id: 1 progress: 0.0*
>
> but the containers that I am getting back is as follows
> [Container: [ContainerId: container_1384381084244_0001_01_000002,
>
> NodeId:
>
> h1:1234, NodeHttpAddress: h1:2, Resource: <memory:1024, vCores:1>,
> Priority: 1, Token: Token { kind: ContainerToken, service: h1:1234 },
>
> ],
>
> Container: [ContainerId: container_1384381084244_0001_01_000003,
>
> NodeId:
>
> h1:1234, NodeHttpAddress: h1:2, Resource: <memory:1024, vCores:1>,
> Priority: 1, Token: Token { kind: ContainerToken, service: h1:1234 },
>
> ]]
>
> I am attaching the test case that I have written along with the mail.
>
> It
>
> uses classes under org.apache.hadoop.yarn.server.resourcemanager
>
> package.
>
> Any pointers would be of great help
>
> Thanks
> Gaurav
>
>
>
>
>
> --
> Alejandro
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>

Mime
View raw message