[codes-ross-users] Help with Storage Simulation

Fri Jul 28 09:26:11 CDT 2017

Hi Harsh,

Parallel conservative mode has a faster performance with models that have a larger look ahead time. The network models typically have a small look ahead value, which causes conservative mode to do a lot of synchronization (that is why it is so slow).

When running in parallel, we typically use optimistic mode (--sync=3). To have the simulations working correctly in optimistic mode, you will have to ensure that whenever you are modifying the model state in the simulation, you will have to provide methods to undo this state change by writing a reverse handler. Majority of our models already implement these reverse handlers. For details on how to write them, you can see _rc functions in the models or look at ROSS documentation:

https://github.com/carothersc/ROSS/wiki/Constructing-the-Model

Thanks,
Misbah
From: Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>
Date: Friday, July 28, 2017 at 9:40 AM
To: Misbah Mubarak <mmubarak at anl.gov<mailto:mmubarak at anl.gov>>
Cc: "codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>" <codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>>
Subject: Re: [codes-ross-users] Help with Storage Simulation

Hi,

I have been working on the codes-storage-server simulation by replacing the checkpoint workload with a darshan workload. Things seem to be working fine, and I am getting the results that I expect but the simulation takes a long time to complete even for moderately sized darshan workloads.

In fact, sequential simulation seems to be running faster than a parallel conservative simulation. For sync=1, I see that the simulation makes much faster progress than it does for sync=2 with 64 or even higher node count.

I have tried other configurations as well but there's no observable speed-up. Any ideas on why that could be happening?

Thanks,
Harsh

On Tue, Jul 11, 2017 at 9:32 AM, Mubarak, Misbah <mmubarak at anl.gov<mailto:mmubarak at anl.gov>> wrote:
Great, thanks for the heads up!

From: Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>
Date: Monday, July 10, 2017 at 4:50 PM
To: Misbah Mubarak <mmubarak at anl.gov<mailto:mmubarak at anl.gov>>

Subject: Re: [codes-ross-users] Help with Storage Simulation

Yes, it works now. Thank you.
I realized what the problem was, the inter and intra group connectivity files paths needed to be absolute in the conf file in my case. It seemed to have a problem picking it up from the environment variable for some reason.

Thanks,
Harsh

On Mon, Jul 10, 2017 at 4:26 PM, Mubarak, Misbah <mmubarak at anl.gov<mailto:mmubarak at anl.gov>> wrote:
FYI I have updated the instructions to point to the right allocation file.  I think the problem was that in your case, there was a mismatch between allocation file and the workloads file.  You can give it a try and let me if that works.

Thanks,
Misbah
From: Misbah Mubarak <mmubarak at anl.gov<mailto:mmubarak at anl.gov>>
Date: Monday, July 10, 2017 at 2:25 PM
To: Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>

Subject: Re: [codes-ross-users] Help with Storage Simulation

Hi Harsh,

The test-checkpoint-dfly-1T config file is actually in the repo:

https://xgitlab.cels.anl.gov/codes/codes-storage-server/blob/master/tests/conf/test-checkpoint-dfly-1T.conf

You will have to adjust the paths for the input network configuration files. The network config files are located in codes at the following path:

codes/src/network-workloads/conf/dragonfly-custom/intra-9K-custom

Thanks,
Misbah
From: Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>
Date: Monday, July 10, 2017 at 3:07 PM
To: Misbah Mubarak <mmubarak at anl.gov<mailto:mmubarak at anl.gov>>
Cc: "codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>" <codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>>
Subject: Re: [codes-ross-users] Help with Storage Simulation

Hi,

Thanks for the help with codes-storage-server.
I was trying to run the client-mul-wklds test on the cluster here using the instructions in the wiki. First I tried with the exact command from the wiki:

./client-mul-wklds --sync=1

--workload-conf-file=/lustre/atlas/scratch/hkhetaw/gen008/CODES-darshan/codes-storage-server/checkpoint-study/workload-files/workload-512.conf

--rank-alloc-file=/lustre/atlas/scratch/hkhetaw/gen008/CODES-darshan/codes-storage-server/checkpoint-study/config-files/ --codes-config=/lustre/atlas/scratch/hkhetaw/gen008/CODES-darshan/codes-storage-server/checkpoint-study/config-files/test-checkpoint-dfly-1T-adap.conf

--lp-io-dir=/lustre/atlas/scratch/hkhetaw/gen008/CODES-darshan/test-dir

--lp-io-use-suffix=1

But the rank-alloc-file is a directory in the wiki instructions, so I expectedly got an error.

Then I tried passing the rank alloc file from the checkpoint-study/workload-files/allocations/contiguous/cont-alloc-8832-2048.conf, but I got this error:

512 instances of workload checkpoint node: 0: error: ../src/networks/model-net/dragonfly-custom.C:630: intra-group file not found

Rank 0 [Mon Jul 10 13:58:50 2017] [c6-1c0s6n2] application called MPI_Abort(MPI_COMM_WORLD, 1) - process 0

_pmiu_daemon(SIGCHLD): [NID 02348] [c6-1c0s6n2] [Mon Jul 10 13:58:50 2017] PE RANK 0 exit signal Aborted

Application 15031648 exit codes: 134

I have set the $HOME_CODES environment variable correctly as well.
Also, the instructions in the wiki specifies the codes-config file as "test-checkpoint-dfly-1T.conf" but there isn't any such file there, so I've tried with both the "test-checkpoint-dfly-1T-adap.conf" and "test-checkpoint-dfly-1T-min.conf", but no luck.

Thanks again for all the help.

Thanks,
Harsh

On Fri, Jul 7, 2017 at 1:58 PM, Mubarak, Misbah <mmubarak at anl.gov<mailto:mmubarak at anl.gov>> wrote:
Hi Harsh,

You can refer to the instructions at the wiki for running the simulation:

https://xgitlab.cels.anl.gov/codes/codes-storage-server/wikis/checkpoint-study

The right config file to use is test-checkpoint-dfly-1T.conf. The one you mentioned is dated so I have removed that to avoid confusion. For a test run, you can reduce the parameter checkpoint_sz in the config (it is currently set to 1TiB which takes some time to complete).

For using the Darshan workload generator, please use Darshan versions 2.x (Darshan 3.0 and up are not supported right now).

Thanks,
Misbah
From: Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>
Date: Friday, July 7, 2017 at 1:36 PM
To: Misbah Mubarak <mmubarak at anl.gov<mailto:mmubarak at anl.gov>>
Cc: "codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>" <codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>>
Subject: Re: [codes-ross-users] Help with Storage Simulation

Hi,

I have been trying to run the tests/test-checkpoint.c simulation as you had suggested, with the intention of building on top of it with darshan workloads. But when I try to run the simulation (without any changes) with:

aprun -n4 ./test-checkpoint --sync=3 --codes-config=/lustre/atlas/scratch/hkhetaw/gen008/CODES-darshan/codes-storage-server/tests/conf/test-checkpoint.conf --lp-io-dir=test-checkpoint-output-ser

I get the following error:

node: 0: error: ../src/util/codes_mapping.c:487: could not find LP with type name "dragonfly_router", did you forget to register the LP?

node: 1: error: ../src/util/codes_mapping.c:487: could not find LP with type name "dragonfly_router", did you forget to register the LP?

node: 2: error: ../src/util/codes_mapping.c:487: could not find LP with type name "dragonfly_router", did you forget to register the LP?

Rank 0 [Fri Jul  7 13:12:24 2017] [c2-1c0s1n2] application called MPI_Abort(MPI_COMM_WORLD, 1) - process 0

I checked in the source code for the test and the LP seems to be registered properly. Any idea why this could be happening?
And would it be the right way to go about it if I used the darshan-workload-generator to issue I/O requests to the storage-model instead of the checkpointing I/O being issued by the test case right now?

Thanks,
Harsh Khetawat

On Fri, Jul 7, 2017 at 9:40 AM, Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>> wrote:
Hi Misbah,

Thank you for the information. This seems perfect for what I want to do. I'll try to set-up a simulation with replays of darshan logs and see how it goes.
Thanks again.

Regards,
Harsh Khetawat

On Thu, Jul 6, 2017 at 4:02 PM, Mubarak, Misbah <mmubarak at anl.gov<mailto:mmubarak at anl.gov>> wrote:
Hi Harsh,

Have you looked at the codes-storage-server model? There is a wiki at:

https://xgitlab.cels.anl.gov/codes/codes-storage-server/wikis/home

The wiki has an example case study here that models storage nodes on the dragonfly network. The clients (compute nodes) interact with the storage nodes via an API that does RDMA style communication. You can find the example on the API usage in tests/test-client-checkpoint.c [Look for the function: send_req_to_store].  The API itself can be found at: codes-store-lp.h.

The repo can be cloned from:
https://xgitlab.cels.anl.gov/codes/codes-storage-server

The CODES storage model makes use of the local-storage-model in CODES. Attached is a diagram showing the sequence of steps for a write operation (read operation is also implemented, you have to specify a flag indicating whether it is a read or a write).

Hope that helps.

Regards,
Misbah
From: <codes-ross-users-bounces at lists.mcs.anl.gov<mailto:codes-ross-users-bounces at lists.mcs.anl.gov>> on behalf of Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>
Date: Thursday, July 6, 2017 at 3:37 PM
To: "codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>" <codes-ross-users at lists.mcs.anl.gov<mailto:codes-ross-users at lists.mcs.anl.gov>>
Cc: Harsh Khetawat <hkhetaw at ncsu.edu<mailto:hkhetaw at ncsu.edu>>
Subject: [codes-ross-users] Help with Storage Simulation

Hi,

I have set-up CODES on my cluster and would like to use it to simulate different storage strategies.

I wanted to use the darshan logs I have collected along with the local-storage-model in CODES to simulate I/O behavior of the application.

My simulation involves not having just one global storage LP, but having: (i) one storage LP per terminal, or (ii) one storage LP per Modelnet_group, or (iii) other strategies.
Would the local-storage-model be the correct choice for this?
Should I write my own storage model for each scenario?
How would I go about incorporating this with the fattree or dragonfly network models?

Thanks for the help.

Regards,
Harsh Khetawat

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/codes-ross-users/attachments/20170728/14ae7470/attachment-0001.html>