I'm working on an application that is critically dependent on the performance of MPI_Alltoall calls with very small messages (less than 4KB) flying among a large number of processes (currently about 200, while the target is in the thousands and more).
I was under the impression, and reading this paper seems to corroborate that impression, that it would be sensible to exploit the hierarchy of a modern PC cluster (as the one I have at my disposal) by segregating processes belonging to a single cluster node within separate communicators (OpenMPI even has a function, from the MPI 3 standard, that does just that) and then, instead of an MPI_Alltoall over MPI_COMM_WORLD, using a sequence MPI_Gather - MPI_Alltoall - MPI_Scatter whereas Gather and Scatter are restricted within those communicators while the Alltoall is over another communicator including one and one only 'gatherer' process per node, in this way exploiting the supposedly faster transfers in internal memory for the Gather and Scatter while hopefully increasing the efficiency of the Alltoall among nodes by having fewer, larger messages passing through the network interfaces (ConnectX InfiniBand NICs by Mellanox, in my case).
I'm trying to verify the assertions of the paper, and if someone is interested I can share my findings, but what I wanted to know is this: perusing the 1.10.x OpenMPI sources, I clearly see a 'hierarch' component among the 'coll' modules, which is also mentioned in the README that makes it seem like such a hierarchical implementation had been pulled in already.
Nevertheless, I was never able to make it work and it seems that it vanished altogether from the 2.x branch (there is none in the 'ompi_info' output).
Is there anyone that succeeded in using it? Can you tell about any improvements, if any, compared to a regular MPI_Alltoall?
Related
I'm new to MPI and I'm trying to understand how MPI (and specifically OpenMPI) work in order to reason about the performance of my system.
I've tried to find resources online to help me understand things a little better, but haven't had much luck. I thought I'd come here.
Right now my question is simple: if I have 3 nodes (1 master, 2 clients) and I issue an MPI_Gather, does the root process handle incoming data sequentially or concurrently? In other words, if processes 1 is the first to make a connection with processes 0, will process 2 have to wait until processes 1 is done sending its data before it can start to send its data?
Thanks!
There are multiple components in Open MPI that implement collective operations and some of them provide multiple algorithms for the implementation of each operation.
What you are most likely interested in is the tuned component of the coll framework as that is what Open MPI uses by default. tuned implements all collectives using point-to-point operations and provides several algorithms for gather:
linear with synchronisation - used when messages are large to mid-size
binomial - used when the number of processes is large or the message size is small
basic linear - used in all other cases
The performance of each algorithm depends strongly on the particular combination of message size and number of ranks, therefore the library comes with a set of heuristics that tries to determine the best algorithm based on the data size and the size of the communicator (as indicated above). There are several mechanisms to override the heuristics and either force a certain algorithm or provide a list of custom algorithm selection rules.
The basic linear algorithm simply has the root loop over all other ranks receiving their messages in sequence. In that case, rank 2 won't be able to send its chunk before rank 1 since the root will first receive the message from rank 1 and only then move on to rank 2.
The linear with synchronisation algorithm splits the chunks into two pieces each. The first pieces are collected in sequence just like in the basic linear algorithm. The second pieces are collected asynchronously using non-blocking receives.
The binomial algorithm arranges the ranks as a binomial tree. The processes at the nodes of the tree receive the chunks from the lower levels and aggregate them into larger chunks that then get passed to the upper levels until they reach the root rank.
You can find the source code of the tuned module in the ompi/mca/coll/tuned folder of the Open MPI source tree. In the development branch, part of the tuned component got promoted to the base implementation of the collective framework and the code for the gather is to be found in ompi/mca/coll/base instead.
Hristo's answer is of course excellent, but I would like to offer a different point of view.
Contrary to your expectation, the question is not simple. It isn't even possible to specifically answer it without knowing more system specifics, as Hristo pointed out. That doesn't mean the question is invalid, but you should start to reason about performance on a different level.
First, consider the complexity of a the gather operation: The total network transfer to the root as well as the memory requirements are linearly growing with the number of processes in the communicator. This naturally limits scalability.
Second, you may assume that your MPI implementation does implement MPI_Gather in the most efficient way possible - better than you could do it by hand. This assumption may very well be wrong, but it is the best starting point to write your program.
Now when you have your program, you should measure and see where time is spent - or wasted. For that you should an MPI performance analysis tools. Now if you have identified that your Gather has a significant impact on performance, you can go ahead and try to optimize that: But to do so, first consider if you can structure your communication conceptually better, e.g. by somehow removing the computation all together or using a clever reduction instead. If you still need to stick to the gather: go ahead and tune your MPI implementation. Afterwards verify that your optimization did indeed improve performance on your specific system.
I want to implement some file io with the routines provided by MPI (in particular Open MPI).
Due to possible limitations of the environment, I wondered, if it is possible to limit the nodes, which are responsible for IO, so that all other nodes are required to perform a hidden mpi_send to this group of processes, to actually write the data. This would be nice in cases, where e.g. the master node is placed on a node with high-performance filesystem and the other nodes have only access to a low-performance filesystem, where the binaries are stored.
Actually, I already found some information, which might be helpful, but I couldn't find further information, how to actually implement these things:
1: There is an info key MPI_IO belonging to the communicator, which tells which ranks provide standard-conforming IO-routines. As this is listed as an environmental inquiry, I don't see, where I could modify this.
2: There is an info key io_nodes_list which seems to belong to file-related info-objects. Unfortunately, the possible values for this key are not documented and Open MPI doesn't seem to implement them in any way. Actually, I can't even get the filename from the info-object which is returned by mpi_file_get_info...
As a workaround, I could imagine two things: On the one hand, I could perform the IO with standard Fortran routines, or on the other hand, create a new communicator, which is responsible for IO. But in both cases, the processes, which are responsible for IO have to check for possible IO from the other processes to perform manual communication and file interaction.
Is there a nice and automatic way to restrict the IO to certain nodes? If yes, how could I implement this?
You explicitly asked about OpenMPI, but there are two MPI-IO implementations in OpenMPI. The old workhorse is ROMIO, the MPI-IO implementation shared among just about every MPI implementation. OpenMPI also has OMPIO, but I don't know a whole lot about tuning that one.
Next, if you want things to happen automatically for you, you'll have to use collective i/o. The independent I/O routines cannot send a message to anyone else -- they are independent and there's no way to know if the other side will be listening.
With those preliminaries out of the way...
You are asking about "i/o aggregaton". There is a bit of information here in the context of another optimization called "deferred open" (and which OMPIO calls Lazy Open)
https://press3.mcs.anl.gov/romio/2003/08/05/deferred-open/
In short, you can definitely say "only these N processes should do I/O", and then the collective I/O library will exchange data and make sure that happens. The optimization was developed some 15-odd years ago for just the situation you proposed: some nodes being better connected to storage than others (as was the case on the old ASCI Red machine, to give you a sense for how old this optimization is...)
I don't know where you got io_nodes_list. You probably want to use the MPI-IO info keys cb_config_list and cb_nodes
So, you've got a cluster with master1, master2, master3, and compute1, compute2, compute3 (or whatever the hostnames actually are). You can do something like this (in c, sorry. I'm not proficient in Fortran):
MPI_Info info;
MPI_File fh;
MPI_Info_create(&info);
MPI_Info_set(info, "cb_config_list", "master1:1,master2:1,master3:1");
MPI_File_open(MPI_COMM_WORLD, filename, MPI_MODE_CREATE|MPI_MODE_WRONLY, info, &fh)
With these hints, MPI_File_write_all will aggregate all the I/O through the MPI processes on master1, master2, and master3. ROMIO won't blow up your memory because it will chunk up the I/O into a smaller working set (specified with the "cb_buffer_size" hint: cranking this up, if you have the memory, is a good way to get better performance).
There is a ton of information about the hints you can set in the ROMIO users guide:
http://www.mcs.anl.gov/research/projects/romio/doc/users-guide/node6.html
I am trying an implementation in MPI where I am invoking multiple slaves (upto 4) on the same machine (localhost) and distributing the computations of my for loop amongst the slaves. MPI is suited for my current application and I cannot take the openMP route.
The variables that are involved are about 50 and all are uni-dimensional arrays.
What would be the best way to send the 50 variables to the master process? Should I send and receive all variables or should I pack them in one 2D array and send this array across to the master?
I am looking for an efficient and computationally inexpensive approach.
Thanks
As so often: it depends. If your individual arrays are sufficiently large, such that latency gets insignificant, it would be fine to send each one individually. Otherwise, it will be better to increase the size of your message by collecting all those arrays into a single one.
If your variables are of different type you could make use of MPI datatypes to describe the layout of your data.
Additionally, if you need to collect this data from multiple processes it might be a good idea to use MPI_Gather or one of its variants.
It might also be, that a viable option in your scenario would be to make use of the one-sided communication facilities offered by MPI.
I have a filtering algorithm that needs to be applied recursively and I am not sure if MapReduce is suitable for this job. W/o giving too much away, I can say that each object that is being filtered is characterized by a collection if ordered list or queue.
The data is not huge, just about 250MB when I export from SQL to
CSV.
The mapping step is simple: the head of the list contains an object that can classify the list as belonging to one of N mapping nodes. the filtration algorithm at each node works on the collection of lists assigned to the node and at the end of the filtration, either a list remains the same as before the filtration or the head of the list is removed.
The reduce function is simple too: all the map jobs' lists are brought together and may have to be written back to disk.
When all the N nodes have returned their output, the mapping step is repeated with this new set of data.
Note: N can be as much as 2000 nodes.
Simple, but it requires perhaps up to a 1000 recursions before the algorithm's termination conditions are met.
My question is would this job be suitable for Hadoop? If not, what are my options?
The main strength of Hadoop is its ability to transparently distribute work on a large number of machines. In order to fully benefit from Hadoop your application has to be characterized, at least by the following three things:
work with large amounts of data (data which is distributed in the cluster of machines) - which would be impossible to store on one machine
be data-parallelizable (i.e. chunks of the original data can be manipulated independently from other chunks)
the problem which the application is trying to solve lends itself nicely to the MapReduce (scatter - gather) model.
It seems that out of these 3, your application has only the last 2 characteristics (with the observation that you are trying to recursively use a scatter - gather procedure - which means a large number of jobs - equal to the recursion depth; see last paragraph why this might not be appropriate for hadoop).
Given the amount of data you're trying to process, I don't see any reason why you wouldn't do it on a single machine, completely in memory. If you think you can benefit from processing that small amount of data in parallel, I would recommend focusing on multicore processing than on distributed data intensive processing. Of course, using the processing power of a networked cluster is tempting but this comes at a cost: mainly the time inefficiency given by the network communication (network being the most contended resource in a hadoop cluster) and by the I/O. In scenarios which are well-fitted to the Hadoop framework these inefficiency can be ignored because of the efficiency gained by distributing the data and the associated work on that data.
As I can see, you need 1000 jobs. The setup and the cleanup of all those jobs would be an unnecessary overhead for your scenario. Also, the overhead of network transfer is not necessary, in my opinion.
Recursive algos are hard in the distributed systems since they can lead to a quick starvation. Any middleware that would work for that needs to support distributed continuations, i.e. the ability to make a "recursive" call without holding the resources (like threads) of the calling side.
GridGain is one product that natively supports distributed continuations.
THe litmus test on distributed continuations: try to develop a naive fibonacci implementation in distributed context using recursive calls. Here's the GridGain's example that implements this using continuations.
Hope it helps.
Q&D, but I suggest you read a comparison of MongoDB and Hadoop:
http://www.osintegrators.com/whitepapers/MongoHadoopWP/index.html
Without knowing more, it's hard to tell. You might want to try both. Post your results if you do!
I have an OpenCL kernel that calculates total force on a particle exerted by other particles in the system, and then another one that integrates the particle position/velocity. I would like to parallelize these kernels across multiple GPUs, basically assigning some amount of particles to each GPU. However, I have to run this kernel multiple times, and the result from each GPU is used on every other. Let me explain that a little further:
Say you have particle 0 on GPU 0, and particle 1 on GPU 1. The force on particle 0 is changed, as is the force on particle 1, and then their positions and velocities are changed accordingly by the integrator. Then, these new positions need to be placed on each GPU (both GPUs need to know where both particle 0 and particle 1 are) and these new positions are used to calculate the forces on each particle in the next step, which is used by the integrator, whose results are used to calculate forces, etc, etc. Essentially, all the buffers need to contain the same information by the time the force calculations roll around.
So, the question is: What is the best way to synchronize buffers across GPUs, given that each GPU has a different buffer? They cannot have a single shared buffer if I want to keep parallelism, as per my last question (though, if there is a way to create a shared buffer and still keep multiple GPUs, I'm all for that). I suspect that copying the results each step will cause more slowdown than it's worth to parallelize the algorithm across GPUs.
I did find this thread, but the answer was not very definitive and applied only to a single buffer across all GPUs. I would like to know, specifically, for Nvidia GPUs (more specifically, the Tesla M2090).
EDIT: Actually, as per this thread on the Khronos forums, a representative from the OpenCL working group says that a single buffer on a shared context does indeed get spread across multiple GPUs, with each one making sure that it has the latest info in memory. However, I'm not seeing that behavior on Nvidia GPUs; when I use watch -n .5 nvidia-smi while my program is running in the background, I see one GPU's memory usage go up for a while, and then go down while another GPU's memory usage goes up. Is there anyone out there that can point me in the right direction with this? Maybe it's just their implementation?
It sounds like you are having implementation trouble.
There's a great presentation from SIGGRAPH that shows a few different ways to utilize multiple GPUs with shared memory. The slides are here.
I imagine that, in your current setup, you have a single context containing multiple devices with multiple command queues. This is probably the right way to go, for what you're doing.
Appendix A of the OpenCL 1.2 specification says that:
OpenCL memory objects, [...] are created using a context and can be shared across multiple command-queues created using the same context.
Further:
The application needs to implement appropriate synchronization across threads on the host processor to ensure that the changes to the state of a shared object [...] happen in the correct order [...] when multiple command-queues in multiple threads are making changes to the state of a shared object.
So it would seem to me that your kernel that calculates particle position and velocity needs to depend on your kernel that calculates the inter-particle forces. It sounds like you already know that.
To put things more in terms of your question:
What is the best way to synchronize buffers across GPUs, given that each GPU has a different buffer?
... I think the answer is "don't have the buffers be separate." Use the same cl_mem object between two devices by having that cl_mem object come from the same context.
As for where the data actually lives... as you pointed out, that's implementation-defined (at least as far as I can tell from the spec). You probably shouldn't worry about where the data is living, and just access the data from both command queues.
I realize this could create some serious performance concerns. Implementations will likely evolve and get better, so if you write your code according to the spec now, it'll probably run better in the future.
Another thing you could try in order to get a better (or a least different) buffer-sharing behavior would be to make the particle data a map.
If it's any help, our setup (a bunch of nodes with dual C2070s) seem to share buffers fairly optimally. Sometimes, the data is kept on only one device, other times it might have the data exist in both places.
All in all, I think the answer here is to do it in the best way the spec provides and hope for the best in terms of implementation.
I hope I was helpful,
Ryan