Request: Option to reduce server data transfer for big workflow in cluster
We've run into a scenario lately where we need to run a very large workflow (huge data in intermediate steps) many times. We can't do this because Galaxy copies all intermediate steps to all notes, which would bog down the servers too much. I asked about something similar before and John mentioned the feature to automatically delete intermediate step data in a workflow once it completed, was coming soon. Is that a feature now? That would help. Ultimately though we can't be copying all this data around to all nodes. The network just isn't good enough, so I have an idea. What if we have an option on the 'run workflow' screen to only run on one node (eliminating the neat Galaxy concurrency ability for that workflow unfortunately)? Then it just propagates the final step data. Or maybe only copy to a couple other nodes, to keep concurrency. If the job errored then in this case I think it should just throw out all the data, or propagate where it stopped. I've been trying to work on implementing this myself but it's taking me a long time. I only just started understanding the pyramid stack, and am putting in the checkbox in the run.mako template. I still need to learn the database schema, message passing, and how jobs are stored, and how to tell condor to only use 1 node, (and more I'm sure) in Galaxy. (I'm drowning) This seems like a really important feature though as Galaxy gains more traction as a research tool for bigger projects that demand working with huge data, and running huge workflows many many times.
Hey Ben, Thanks for the e-mail. I did not promise anything was coming soon, I only said people were working on parts of it. It is not a feature yet unfortunately - multiple people including myself are thinking about various parts of this problem though. I would like to respond, but I am trying to understand this line: "We can't do this because Galaxy copies all intermediate steps to all no(d)es, which would bog down the servers too much." Can you describe how you are doing this staging for me? Is data currently being copied around to all the nodes, if so how are you doing that? Or are you trying to say that Galaxy requires the data to be available on all of the nodes? -John On Tue, Dec 17, 2013 at 11:15 AM, Ben Gift <corn8bit2@gmail.com> wrote:
We've run into a scenario lately where we need to run a very large workflow (huge data in intermediate steps) many times. We can't do this because Galaxy copies all intermediate steps to all notes, which would bog down the servers too much.
I asked about something similar before and John mentioned the feature to automatically delete intermediate step data in a workflow once it completed, was coming soon. Is that a feature now? That would help.
Ultimately though we can't be copying all this data around to all nodes. The network just isn't good enough, so I have an idea.
What if we have an option on the 'run workflow' screen to only run on one node (eliminating the neat Galaxy concurrency ability for that workflow unfortunately)? Then it just propagates the final step data.
Or maybe only copy to a couple other nodes, to keep concurrency.
If the job errored then in this case I think it should just throw out all the data, or propagate where it stopped.
I've been trying to work on implementing this myself but it's taking me a long time. I only just started understanding the pyramid stack, and am putting in the checkbox in the run.mako template. I still need to learn the database schema, message passing, and how jobs are stored, and how to tell condor to only use 1 node, (and more I'm sure) in Galaxy. (I'm drowning)
This seems like a really important feature though as Galaxy gains more traction as a research tool for bigger projects that demand working with huge data, and running huge workflows many many times.
___________________________________________________________ Please keep all replies on the list by using "reply all" in your mail client. To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/
To search Galaxy mailing lists use the unified search at: http://galaxyproject.org/search/mailinglists/
Hi John, thanks for the reply. Yes, I mean Galaxy's default behavior of keeping all the data on all nodes of our condor cluster. So for instance if I run a job, then the output of that job is copied to every node in the cluster. Is this not the normal behavior? On Tue, Dec 17, 2013 at 9:42 AM, John Chilton <chilton@msi.umn.edu> wrote:
Hey Ben,
Thanks for the e-mail. I did not promise anything was coming soon, I only said people were working on parts of it. It is not a feature yet unfortunately - multiple people including myself are thinking about various parts of this problem though.
I would like to respond, but I am trying to understand this line: "We can't do this because Galaxy copies all intermediate steps to all no(d)es, which would bog down the servers too much."
Can you describe how you are doing this staging for me? Is data currently being copied around to all the nodes, if so how are you doing that? Or are you trying to say that Galaxy requires the data to be available on all of the nodes?
-John
We've run into a scenario lately where we need to run a very large workflow (huge data in intermediate steps) many times. We can't do this because Galaxy copies all intermediate steps to all notes, which would bog down
servers too much.
I asked about something similar before and John mentioned the feature to automatically delete intermediate step data in a workflow once it completed, was coming soon. Is that a feature now? That would help.
Ultimately though we can't be copying all this data around to all nodes. The network just isn't good enough, so I have an idea.
What if we have an option on the 'run workflow' screen to only run on one node (eliminating the neat Galaxy concurrency ability for that workflow unfortunately)? Then it just propagates the final step data.
Or maybe only copy to a couple other nodes, to keep concurrency.
If the job errored then in this case I think it should just throw out all the data, or propagate where it stopped.
I've been trying to work on implementing this myself but it's taking me a long time. I only just started understanding the pyramid stack, and am putting in the checkbox in the run.mako template. I still need to learn
On Tue, Dec 17, 2013 at 11:15 AM, Ben Gift <corn8bit2@gmail.com> wrote: the the
database schema, message passing, and how jobs are stored, and how to tell condor to only use 1 node, (and more I'm sure) in Galaxy. (I'm drowning)
This seems like a really important feature though as Galaxy gains more traction as a research tool for bigger projects that demand working with huge data, and running huge workflows many many times.
___________________________________________________________ Please keep all replies on the list by using "reply all" in your mail client. To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/
To search Galaxy mailing lists use the unified search at: http://galaxyproject.org/search/mailinglists/
Hey Ben, Hmmm... I don't think Galaxy is doing that - not directly anyway. Unless I am mistaken, Galaxy will put the file in one location on the web server node or VM. Typically, this location is on a filesystem that is shared between the web server and a cluster's compute nodes. So I wouldn't describe that as Galaxy copying the data to all of the compute nodes. If your cluster doesn't have a shared file system and to get around this someone has configured the Galaxy data to be synced across all nodes - that would be copying the data to all nodes but I don't think Galaxy is doing that. Alternatively, you might have condor configured to copy the galaxy data to the remote nodes in such a way that everytime a Galaxy job is run on a node, all the data is copied to that node? Does that make sense? So I still don't entirely understand your setup, but my advice is pretty general - for now you may want to solve this problem at the condor level. I am assuming this is a general purpose condor cluster and not setup explicitly for Galaxy? Lets say you have 200 nodes in your condor cluster and they cannot all mount the Galaxy filesystem - because it would overload the file server being used by Galaxy. I think you could setup a FileSystemDomain at the condor level that just 10 of nodes belonged to say (these 10 nodes can continue to run anything in general but Galaxy will only submit to these) This filesystemdomain could have a name like galaxy.example.com if example.com is your default filesystemdomain. Then you can setup the Galaxy condor runner with a requirement such that "FileSystemDomain=galaxy.example.com" and Galaxy jobs will only run on these 10 nodes then. Having 10 nodes mount a file server is much more manageable than 200. -John On Tue, Dec 17, 2013 at 11:52 AM, Ben Gift <corn8bit2@gmail.com> wrote:
Hi John, thanks for the reply.
Yes, I mean Galaxy's default behavior of keeping all the data on all nodes of our condor cluster. So for instance if I run a job, then the output of that job is copied to every node in the cluster. Is this not the normal behavior?
On Tue, Dec 17, 2013 at 9:42 AM, John Chilton <chilton@msi.umn.edu> wrote:
Hey Ben,
Thanks for the e-mail. I did not promise anything was coming soon, I only said people were working on parts of it. It is not a feature yet unfortunately - multiple people including myself are thinking about various parts of this problem though.
I would like to respond, but I am trying to understand this line: "We can't do this because Galaxy copies all intermediate steps to all no(d)es, which would bog down the servers too much."
Can you describe how you are doing this staging for me? Is data currently being copied around to all the nodes, if so how are you doing that? Or are you trying to say that Galaxy requires the data to be available on all of the nodes?
-John
On Tue, Dec 17, 2013 at 11:15 AM, Ben Gift <corn8bit2@gmail.com> wrote:
We've run into a scenario lately where we need to run a very large workflow (huge data in intermediate steps) many times. We can't do this because Galaxy copies all intermediate steps to all notes, which would bog down the servers too much.
I asked about something similar before and John mentioned the feature to automatically delete intermediate step data in a workflow once it completed, was coming soon. Is that a feature now? That would help.
Ultimately though we can't be copying all this data around to all nodes. The network just isn't good enough, so I have an idea.
What if we have an option on the 'run workflow' screen to only run on one node (eliminating the neat Galaxy concurrency ability for that workflow unfortunately)? Then it just propagates the final step data.
Or maybe only copy to a couple other nodes, to keep concurrency.
If the job errored then in this case I think it should just throw out all the data, or propagate where it stopped.
I've been trying to work on implementing this myself but it's taking me a long time. I only just started understanding the pyramid stack, and am putting in the checkbox in the run.mako template. I still need to learn the database schema, message passing, and how jobs are stored, and how to tell condor to only use 1 node, (and more I'm sure) in Galaxy. (I'm drowning)
This seems like a really important feature though as Galaxy gains more traction as a research tool for bigger projects that demand working with huge data, and running huge workflows many many times.
___________________________________________________________ Please keep all replies on the list by using "reply all" in your mail client. To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/
To search Galaxy mailing lists use the unified search at: http://galaxyproject.org/search/mailinglists/
This does mostly make sense, and is very illuminating. I appreciate all the help and I'm sorry that I'm so new to this. I'm not sure I fully understand though. Do you mean that I could have a main Galaxy install setup for the 200 nodes, for general purpose use with a shared file system, and a specialized setup that only uses 10 with its own shared file system so that it doesn't bog down the main one? And to use the specialized install when running huge workflows? On Tue, Dec 17, 2013 at 10:11 AM, John Chilton <chilton@msi.umn.edu> wrote:
Hey Ben,
Hmmm... I don't think Galaxy is doing that - not directly anyway. Unless I am mistaken, Galaxy will put the file in one location on the web server node or VM. Typically, this location is on a filesystem that is shared between the web server and a cluster's compute nodes. So I wouldn't describe that as Galaxy copying the data to all of the compute nodes. If your cluster doesn't have a shared file system and to get around this someone has configured the Galaxy data to be synced across all nodes - that would be copying the data to all nodes but I don't think Galaxy is doing that. Alternatively, you might have condor configured to copy the galaxy data to the remote nodes in such a way that everytime a Galaxy job is run on a node, all the data is copied to that node?
Does that make sense?
So I still don't entirely understand your setup, but my advice is pretty general - for now you may want to solve this problem at the condor level. I am assuming this is a general purpose condor cluster and not setup explicitly for Galaxy? Lets say you have 200 nodes in your condor cluster and they cannot all mount the Galaxy filesystem - because it would overload the file server being used by Galaxy. I think you could setup a FileSystemDomain at the condor level that just 10 of nodes belonged to say (these 10 nodes can continue to run anything in general but Galaxy will only submit to these) This filesystemdomain could have a name like galaxy.example.com if example.com is your default filesystemdomain. Then you can setup the Galaxy condor runner with a requirement such that "FileSystemDomain=galaxy.example.com" and Galaxy jobs will only run on these 10 nodes then. Having 10 nodes mount a file server is much more manageable than 200.
-John
On Tue, Dec 17, 2013 at 11:52 AM, Ben Gift <corn8bit2@gmail.com> wrote:
Hi John, thanks for the reply.
Yes, I mean Galaxy's default behavior of keeping all the data on all nodes of our condor cluster. So for instance if I run a job, then the output of that job is copied to every node in the cluster. Is this not the normal behavior?
On Tue, Dec 17, 2013 at 9:42 AM, John Chilton <chilton@msi.umn.edu> wrote:
Hey Ben,
Thanks for the e-mail. I did not promise anything was coming soon, I only said people were working on parts of it. It is not a feature yet unfortunately - multiple people including myself are thinking about various parts of this problem though.
I would like to respond, but I am trying to understand this line: "We can't do this because Galaxy copies all intermediate steps to all no(d)es, which would bog down the servers too much."
Can you describe how you are doing this staging for me? Is data currently being copied around to all the nodes, if so how are you doing that? Or are you trying to say that Galaxy requires the data to be available on all of the nodes?
-John
On Tue, Dec 17, 2013 at 11:15 AM, Ben Gift <corn8bit2@gmail.com> wrote:
We've run into a scenario lately where we need to run a very large workflow (huge data in intermediate steps) many times. We can't do this because Galaxy copies all intermediate steps to all notes, which would bog
down
the servers too much.
I asked about something similar before and John mentioned the feature to automatically delete intermediate step data in a workflow once it completed, was coming soon. Is that a feature now? That would help.
Ultimately though we can't be copying all this data around to all nodes. The network just isn't good enough, so I have an idea.
What if we have an option on the 'run workflow' screen to only run on one node (eliminating the neat Galaxy concurrency ability for that workflow unfortunately)? Then it just propagates the final step data.
Or maybe only copy to a couple other nodes, to keep concurrency.
If the job errored then in this case I think it should just throw out all the data, or propagate where it stopped.
I've been trying to work on implementing this myself but it's taking me a long time. I only just started understanding the pyramid stack, and am putting in the checkbox in the run.mako template. I still need to learn the database schema, message passing, and how jobs are stored, and how to tell condor to only use 1 node, (and more I'm sure) in Galaxy. (I'm drowning)
This seems like a really important feature though as Galaxy gains more traction as a research tool for bigger projects that demand working with huge data, and running huge workflows many many times.
___________________________________________________________ Please keep all replies on the list by using "reply all" in your mail client. To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/
To search Galaxy mailing lists use the unified search at: http://galaxyproject.org/search/mailinglists/
How do you have it set up on the main public galaxy install? I imagine that people run enough big jobs that there there is enormous use of your shared file system. How did you scale that to so many nodes without bogging down the file system with large dataset transfers? It seems that for now the solution of having a second Galaxy instance will work well, thank you very much John :) . But I'm still interested in a more permanent scaled solution. After reading up more on our shared file system it still seems like heavy traffic is bad, so could my initial idea still be good?
File system performance varies wildly between storage architectures. There are storage server setups that can easily scale to orders of magnitude beyond the compute that backs usegalaxy.org - suffice to say we are currently bound by the number of cores we have available and not by IO/network performance. (Nate may elaborate on the specifics of the public servers setup, but I am not sure it is useful to you unless you have hundreds or thousands (or millions) of dollars to spend on new storage and network hardware :) ). Also my idea was not a second Galaxy instance - sorry I did not make that clearer. It was to restrict your current Galaxy instance to a smaller portion of your cluster. If your cluster is completely dedicated to Galaxy however this idea doesn't make sense, but if this is a shared condor cluster used for other things (besides Galaxy) it could make sense. Sorry have not been more helpful. -John On Tue, Dec 17, 2013 at 6:12 PM, Ben Gift <corn8bit2@gmail.com> wrote:
How do you have it set up on the main public galaxy install? I imagine that people run enough big jobs that there there is enormous use of your shared file system. How did you scale that to so many nodes without bogging down the file system with large dataset transfers?
It seems that for now the solution of having a second Galaxy instance will work well, thank you very much John :) . But I'm still interested in a more permanent scaled solution. After reading up more on our shared file system it still seems like heavy traffic is bad, so could my initial idea still be good?
You've been extremely helpful, I appreciate it. So we went ahead and decided that we need this feature. We're planning to have a lot of people running huge pipelines, ones that work best on one node, and there's no reason to do all this writing to any shared file system when it works best on one node using /tmp/ for intermediate step data. So I've been working on that. So far I've made the checkbox for using one node (in run.mako). In workflow.py I catch this and set a new variable in each step of the workflow called use_one_node, if checkbox is checked. Now I'm trying to find where jobs are run, so that I can put the logic in for getting a node to run on, and setting that as a variable on each step. Could you point me in the direction of the files/classes associated with running the history's jobs, and getting nodes (or sending jobs to condor?)? Thanks, and I'll be sure to push this upstream after it's done if you'd like it. Maybe as something you can turn on from the universal_wsgi.
Hi Ben, The job running code is in lib/galaxy/jobs/. Galaxy jobs get a "wrapper" which include the start/finish methods, that's in __init__.py. handler.py is what dispatches jobs out to various runner plugins, finds new jobs to run, and generally controls the operation. runners/*.py are the individual DRM plugins. This is an interesting solution and I'd like to see the implementation. --nate On Thu, Dec 19, 2013 at 6:33 PM, Ben Gift <corn8bit2@gmail.com> wrote:
You've been extremely helpful, I appreciate it.
So we went ahead and decided that we need this feature. We're planning to have a lot of people running huge pipelines, ones that work best on one node, and there's no reason to do all this writing to any shared file system when it works best on one node using /tmp/ for intermediate step data. So I've been working on that.
So far I've made the checkbox for using one node (in run.mako). In workflow.py I catch this and set a new variable in each step of the workflow called use_one_node, if checkbox is checked.
Now I'm trying to find where jobs are run, so that I can put the logic in for getting a node to run on, and setting that as a variable on each step. Could you point me in the direction of the files/classes associated with running the history's jobs, and getting nodes (or sending jobs to condor?)?
Thanks, and I'll be sure to push this upstream after it's done if you'd like it. Maybe as something you can turn on from the universal_wsgi.
___________________________________________________________ Please keep all replies on the list by using "reply all" in your mail client. To manage your subscriptions to this and other Galaxy lists, please use the interface at: http://lists.bx.psu.edu/
To search Galaxy mailing lists use the unified search at: http://galaxyproject.org/search/mailinglists/
participants (3)
-
Ben Gift -
John Chilton -
Nate Coraor