Re: [Scalablecr-discuss] local storage reliability experience?

[Kathryn M. <ka...@ll...>]

Hi Andrew,

>> I am not aware of it, although you might be able to get it to work out of
>> the box depending on how you plan to use the files. What I mean is, if it's
>> OK with your application that SCR treats those files like checkpoints
>> (meaning possibly deletes older ones when new ones are written), then it
>> should be fine. Otherwise, you would need to copy each file out to your
>> parallel file system, which would defeat the purpose of using SCR.
>> 
>> That said, we are planning to add support for ordinary files into SCR in the
>> future. It would be interesting to hear what your use case is and how SCR
>> could best support that.
> 
> Thanks, Kathryn. My interest is indirect. I don't have an app to use
> SCR with, but I am a operating system, storage, and network researcher
> that is interested in how file systems can be better designed for
> users. I'm glad to hear that SCR will support ordinary files because
> it seems like there's a good opportunity for it to take on the use
> cases that traditional parallel file systems have difficulty with.

That was our thinking too.
> 
> What will the support for ordinary files look like? Will those files
> persist on SSD between jobs?

We're not sure what the support will look like at this point -- this is still in the early design stages. However, we imagine an API where applications can tell SCR what type of file is being handled and what to do with it, e.g. keep it in in-system storage, or only cache it in in-system storage, move it to the parallel file system in the background, and then delete the cached copy. 

As to whether files could persist on SSDs between jobs, I imagine that is would depend on the resource management policies. For example, if the same user/app doesn't get the same set of nodes in the next run you wouldn't want the files there necessarily -- for space reasons and maybe privacy/security. 

I'm glad to know of your interest! 

Kathryn 
> 
> Thanks again,
> 
> Andrew

_________________________________________________________________
Kathryn Mohror, ka...@ll..., http://scalability.llnl.gov/
Scalability Team @ Lawrence Livermore National Laboratory, Livermore, CA, USA