Re: POSIX file system tests.


Pittman, Ashley M
 

 

I’m not fully aware of the details but I think it’s more in the OFI layer so it’ll affect everything equally.  It’s something we’re investigating currently.

 

Ashley,

 

From: <daos@daos.groups.io> on behalf of Colin Ngam <colin.ngam@...>
Reply to: "daos@daos.groups.io" <daos@daos.groups.io>
Date: Tuesday, 21 April 2020 at 02:27
To: "daos@daos.groups.io" <daos@daos.groups.io>
Subject: Re: [daos] POSIX file system tests.

 

Hi Ashley,

 

The multi-threading anomalies .. is that in the dfs layer or dfuse or fuse?

 

Thanks.

 

Colin

 

From: <daos@daos.groups.io> on behalf of "Pittman, Ashley M" <ashley.m.pittman@...>
Reply-To: "daos@daos.groups.io" <daos@daos.groups.io>
Date: Monday, April 20, 2020 at 4:59 PM
To: "daos@daos.groups.io" <daos@daos.groups.io>
Subject: Re: [daos] POSIX file system tests.

 

 

Colin

 

In terms of what goes over the network, and therefore peak bandwidth to a node dfs_write, IL (interception library) write and fuse write should be the same, for read dfs_read and fuse are currently slightly more efficient than IL read because of the file size check but we are working to remove this difference.  That said, using fuse will certainly reduce the bandwidth available via a single stream as there’s a lot more shuffling of data within a node so I’d expect this to see lower bandwidth – but to compensate for this we expect the IL to be used which should be comparable with dfs overall.  There are some anomalies with threading performance that can cause issues currently, and we’re also working on fuse to make read/write asynchronous within dfuse which would reduce CPU overhead whilst waiting for I/O so this should close the gap for some workloads.

 

In terms of setting expectations if you are using fuse then I’d expect you to also be using the interception library if at all possible, and if your application supports dfs directly then I’d expect you to use that over fuse in most cases.

 

As Mohamad said there’s the option of enabling caching in dfuse as well which can change the dynamic, but we strongly recommend using a container per client node if this is the case, this would work well for multiple processes on a node reading the same data for example.

 

Ashley,

 

From: <daos@daos.groups.io> on behalf of Colin Ngam <colin.ngam@...>
Reply to: "daos@daos.groups.io" <daos@daos.groups.io>
Date: Monday, 20 April 2020 at 18:10
To: "daos@daos.groups.io" <daos@daos.groups.io>
Subject: Re: [daos] POSIX file system tests.

 

Hi Mohamad,

 

What’s the expectation with respect to performance for dfuse (not direct interface into dfs API). For example, ior over dfuse (dfuse and defuse_hl) versus ior over daos dfs or using libioil for intercept?

 

Thanks for your response. Just setting expectations.

 

Thanks.

 

Colin

 

From: <daos@daos.groups.io> on behalf of "Chaarawi, Mohamad" <mohamad.chaarawi@...>
Reply-To: "daos@daos.groups.io" <daos@daos.groups.io>
Date: Monday, April 20, 2020 at 11:58 AM
To: "daos@daos.groups.io" <daos@daos.groups.io>
Subject: Re: [daos] POSIX file system tests.

 

Hi Colin,

 

There is a small DFS specific test suite that is run part of daos_test (option -F). We are in the process of expanding this test more. Ashley probably can comment if there is any specific suite he runs over dfuse now.

 

But most of the regression tests in CI are done over other middleware/benchmarks:

-          mdtest, ior, fio runs with different configurations.

-          MPI-IO  tests (several suites exercised there).

-          HDF5 using the MPI-IO VFD.

 

We have run other tests like simul, but most of those have not been integrated into our CI yet.

 

Thanks,

Mohamad

 

From: <daos@daos.groups.io> on behalf of Colin Ngam <colin.ngam@...>
Reply-To: "daos@daos.groups.io" <daos@daos.groups.io>
Date: Monday, April 20, 2020 at 10:21 AM
To: "daos@daos.groups.io" <daos@daos.groups.io>
Subject: [daos] POSIX file system tests.

 

Greetings,

 

Just a quick question.

 

What regression tests is executed to test DAOS Posix functionality.

 

Thanks.

 

Colin

 

---------------------------------------------------------------------
Intel Corporation (UK) Limited
Registered No. 1134945 (England)
Registered Office: Pipers Way, Swindon SN3 1RJ
VAT No: 860 2173 47

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.

---------------------------------------------------------------------
Intel Corporation (UK) Limited
Registered No. 1134945 (England)
Registered Office: Pipers Way, Swindon SN3 1RJ
VAT No: 860 2173 47

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.

Join daos@daos.groups.io to automatically receive all group messages.