[Swift-devel] swift-on-ec2

[Kate Keahey]

Ian Foster wrote:
> Kate:
> 
> If we configure the virtual cluster with a full LRM, as you propose (and it seems have already done--great work!), then we can use this to start Falkon executors--as we do today on regular clusters. So it seems to me that we have all we need. How about you and Ioan spend your time on Thursday running something on EC2, to make sure it sorks?

As I suggest below, I think it would be easiest if we could deploy and 
debug a small static cluster locally first, and we can probably give it 
a shot tomorrow. We still don't have access to the Xen nodes on TeraPort 
(although hopefully that might change by tomorrow) but I asked Rick to 
rebuild a couple of nodes at ANL and he did, I think for a test that 
should give us enough resources to play with.

At the same time -- if there are multiple ways of doing this, and 
perhaps better ways than simply using a virtual cluster, we should 
discuss them now. It is not completely clear to me what the relationship 
between Falkon and Swift is, and what the specific objectives are (other 
than that dynamically provisioning resources is required). It looks at 
this point like the objectives probably overlap with what Ioan, Borja 
and I wanted to talk about (which I thought was a separate project, but 
am thrilled to find out is related) so how about we come up with a 
design tomorrow and post the notes on this list (is this a good venue 
for that?) and then others can shoot them down.

> Regarding choice of LRM: have you looked at SGE? That is what quite a few others seem to be using.

Yes, we have. We also collaborate with others who do, as well as with 
Sun... As you may remember, Borja did the scheduling work for his thesis 
in the context of SGE. Last time we talked though, Torque was the 
scheduler of choice for the virtual cluster LRM so we used that.

The usage of SGE you are referring to above -- is this in the context of 
virtualization projects, or as LRM for various Falkon-related applications?

> 
> Ian
> 
> 
> 
> Sent via BlackBerry from T-Mobile  
> 
> -----Original Message-----
> From: Kate Keahey <keahey at mcs.anl.gov>
> Date: Tue, 15 May 2007 23:28:07 
> To:iraicu at cs.uchicago.edu
> Cc:swift-devel at ci.uchicago.edu
> Subject: Re: [Swift-devel] swift-on-ec2
> 
> First -- this is a very useful discussion, would it be possible to see 
> all of it. We need to understand the requirements and trade-offs in some 
> detail to figure out the best way to make this work. I see a few 
> different interaction threads somewhat mixed up here though so just to 
> make sure we are all on the same wavelength, here is some context.
> 
> Ian and I have been talking on and off about providing a workspace 
> service implementation with EC2 backend. The benefit for that would be 
> that users could deploy the same VMs using the same interface to either 
> TeraPort or EC2 or yet another resource provider. The workspace service 
> would also provide some features on top of EC2 (translating between PKI 
> credentials and Amazon's paying accounts, contextualization as needed to 
> make deployment dynamic). One application of interest for this was 
> Swift. Last time we chatted about this though was in the context of 
> using EC2 to provide a production platform for STAR runs (since 
> virtualizing enough TeraPort to provide a production platform is taking 
> a long time). This is what Tim and I are trying to make happen now.
> 
> Since there was also interest in running Swift in VMs, Mike, Tibi and I 
> met around February/March and agreed that a reasonable way to proceed 
> will be for us to stand up a base virtual cluster somewhere locally 
> (e.g., a static deployment on TeraPort) so that they can finish the 
> configuration according to their needs, look at performance, figure out 
> the best way to interact with it, and make sure that there are no 
> VM-induced gotchas. All of this will be much easier to assess locally 
> and on a static deployment. Then we'd make sure the cluster is 
> dynamically deployable using the workspace service (on EC2 or whatever 
> other provider). During the meeting (and over following emails) we 
> agreed that the required "base cluster" would be configured with 
> GRAM/Torque on the headnode plus a number of worker nodes, plus root 
> privileges. We configured this cluster and it is ready to deploy. Are 
> you saying now that in fact something different is needed?
> 
> As Ian says, Borja and I were planning to meet with Ioan on Thursday to 
> discuss interaction between Falkon and the workspace service (not 
> necessarily/exclusively in the EC2 context). I don't completely 
> understand the relationship between swift and falkon -- are there 
> specific applications or scenarios that you are trying to target in this 
> exercise?
> 
> Ioan Raicu wrote:
>> Hi,
>> See below:
>>
>> Tim Freeman wrote:
>>> On Tue, 15 May 2007 16:20:03 +0000 (GMT)
>>> Ben Clifford <benc at hawaga.org.uk> wrote:
>>>
>>>  
>>>> Ian asked about this elsewhere, but its perhaps interesting for 
>>>> swift-devel people to look at the questions too.
>>>>
>>>> On Tue, 15 May 2007, Ian Foster wrote:
>>>>
>>>>    
>>>>> Dear All:
>>>>>       
>>>>                                                                                 
>>>>    
>>>>> I asked Kate if she and Tim could look into creating VM images that 
>>>>> would allow us to run Swift applications on Amazon EC2. I think Kate 
>>>>> is meeting with Ioan about this on Thursday (?).
>>>>>       
>>>>                                                                                 
>>>>    
>>>>> One issue that I thought would be good to discuss is what we'd want 
>>>>> in that VM image. Perhaps this is obvious to the rest of you, but it 
>>>>> isn't to me. A few thoughts:
>>>>>       * I'm assuming that we want to run "workers" on EC2 nodes, and 
>>>>> have the
>>>>> "task dispatch" logic run on some external frontend system outside EC2.
>>>>>       * I would think that we want to use Falkon to do the task 
>>>>> dispatch. If so,
>>>>> we need a Falkon executor on each VM, configured to check in with 
>>>>> the Falkon
>>>>> dispatcher. (Alternatively, we could use, say, SGE: in that case, we 
>>>>> would
>>>>> want an SGE agent.)
>>>>>       *  We need a way of getting data to and from the worker nodes. 
>>>>> Do we want to
>>>>> run a file system across the EC2 nodes and the external frontend 
>>>>> node? That
>>>>> seems rather inefficient. Other options?
>>>>>       * Should we preload the application code on each EC2 node?
>>>>>       
>>>> Here's a couple of approaches:
>>>>
>>>>  1) swift regards all the EC2 nodes that we are paying for as a 
>>>> single     site.
>>>>
>>>> Something like falkon handles all the task dispatch and worker node 
>>>> management. I don't know what that looks like at the moment in 
>>>> Falkon, but the interface for Swift to send jobs into Falkon sounds 
>>>> pretty straightforward and shouldn't need changing.
>>>>     
>>> So if I understand, here there would be no gateway+LRM but each EC2 
>>> node +
>>> Falkon would need a port open to receive tasks?  Or does each node 
>>> pull down
>>> instructions OK from behind a firewall?
>>>   
>> Falkon supports both polling and notifications.  To use notifications, 
>> there needs to be an open port on the worker :(
>>> Is there a latency problem with running each node as an indepdent task
>>> receiver with the dispatcher off-site from EC2?  I would think it 
>>> would be
>>> better to put the queues to fill with tasks on EC2 so it can more 
>>> quickly get
>>> the task going when a node is done with a previous task (I may be 
>>> missing some
>>> nuances here with respect to Falkon, don't know much about this yet!).   
>> We have run the Falkon dispatcher at UChicago and workers at ANL without 
>> any issues, so it can easily tolerate a few ms of latency.  We haven't 
>> tried it across 10s of ms of latency links, but my instinct says that if 
>> you have enough workers, you might be able to hide the latency.  We'd 
>> have to experiment with it to see what happens.  We could potentially do 
>> some experiments between SDSC and ANL over a 50+ ms link, and see what 
>> difference in throughputs we get.
>>
>> Ioan
>>> If a gateway node is desired, this option sounds a lot like the GRAM+LRM
>>> situation we use on VMs with the workspace service and will soon use 
>>> on EC2 via
>>> the workspace EC2 gateway we're implementing.  Start up one gateway 
>>> node and
>>> then add compute nodes which dynamically join the pool, they are 
>>> pointed to the
>>> GRAM node.
>>>
>>>  
>>>> All the nodes in a site are required by our site model to have a 
>>>> shared filesystem - we've talked about removing it, but I think that 
>>>> is still the case and if so, isn't going to change soon.     
>>> Setting up a shared filesystem in this environment is akin to setting 
>>> up the
>>> compute nodes to join an LRM pool.  The VMs can communicate over the 
>>> private
>>> network at EC2, you can instruct EC2 to let all the nodes be open to 
>>> each other
>>> (while simultaneously keeping a separate policy of blocking ports from 
>>> being
>>> open from the internet and other people's EC2 nodes).  The 
>>> non-file-serving
>>> nodes would simply need to know the private address of the filesystem 
>>> server
>>> (unless you are using a fancier network file system than NFS-style ones).
>>> For background: every VM on EC2 currently gets a public address -- 
>>> NAT'd to a
>>> private address which is actually what the VM's one NIC is configured 
>>> with.
>>> There is a facility to open/forward specific network ports on the public
>>> address to each VM either via a group policy or on a VM by VM basis.
>>>
>>> [...]  
>>>> Amazon also has a storage cloud, alongside its compute cloud. I know 
>>>> very little about that and have never thought about how it would fit 
>>>> into the above (if at all). Maybe someone else knows more.
>>>>     
>>> A VM template on EC2 is called an AMI which stands for Amazon Machine 
>>> Image.
>>> This is just a packaging thing but what it mostly means is that the VM is
>>> stored on S3 and also registered into the EC2 system.
>>>
>>> When starting an instance of an AMI, the file is copied from S3 to the
>>> hypervisor node (what we call propagation in the workspace service).  
>>> After it
>>> is used, this file is deleted (an option in the workspace service but 
>>> there is
>>> also an option to save it back with any changes). 
>>> So the VMs are stored in S3 but anything that happens on them after being
>>> started is lost unless you manually do something about it.
>>>
>>> As for free scratch space, you get a good amount per node, 140G.  But 
>>> the node
>>> could go down at any moment just like a physical resource.
>>>
>>> To harness S3 for safely persisting any data (or if you need more 
>>> space) you
>>> would need to actually run S3 clients on the VMs when they are run on 
>>> EC2.  You
>>> could alternatively mirror data between nodes assuming that all would 
>>> not go
>>> down at once.
>>> The good thing is that you do not pay transfer costs between S3 and 
>>> EC2 if you
>>> chose to use S3 for big storage, you would only pay the "housing fees" 
>>> so to
>>> speak.
>>> Tim
>>> _______________________________________________
>>> Swift-devel mailing list
>>> Swift-devel at ci.uchicago.edu
>>> http://mail.ci.uchicago.edu/mailman/listinfo/swift-devel
>>>
>>>   
> 

-- 

Kate Keahey,
Mathematics & CS Division, Argonne National Laboratory
Computation Institute, University of Chicago