What is your question?
I am trying to implement a metric which needs access to whole data. So instead of updating the metric in *_step() methods, I am trying to collect the outputs in the *_epoch_end() methods. However, the outputs contain only the output of the partition of the data each device gets. Basically if there are n devices, then each device is getting 1/n of the total outputs.
What's your environment?
OS: ubuntu
Packaging: conda
Version [1.0.4
Pytorch: 1.6.0
See the pytorch-lightningmanual. I think you are looking for training_step_end/validation_step_end (assuming you are using DP/DDP2).
...So, when Lightning calls any of the training_step, validation_step, test_step you will only be operating on one of those pieces. (...) For most metrics, this doesn’t really matter. However, if you want to add something to your computational graph (like softmax) using all batch parts you can use the training_step_end step.
When using the DDP backend, there's a separate process running for every GPU. They don't have access to each other's data, but there are a few special operations (reduce, all_reduce, gather, all_gather) that make the processes synchronize. When you use such operations on a tensor, the processes will wait for each other to reach the same point and combine their values in some way, for example take the sum from every process.
In theory it's possible to gather all data from all processes and then calculate the metric in one process, but this is slow and prone to problems, so you want to minimize the data that you transfer. The easiest approach is to calculate the metric in pieces and then for example take the average. self.log() calls will do this automatically when you use sync_dist=True.
If you don't want to take the average over the GPU processes, it's also possible to update some state variables at each step, and after the epoch synchronize the state variables and calculate your metric from those values. The recommended way is to create a class that uses the Metrics API, which recently moved from PyTorch Lightning to the TorchMetrics project.
If it's not enough to store a set of state variables, you can try to make your metric gather all data from all the processes. Derive your own metric from the Metric base class, overriding the update() and compute() methods. Use add_state("data", default=[], dist_reduce_fx="cat") to create a list where you collect the data that you need for calculating the metric. dist_reduce_fx="cat" will cause the data from different processes to be combined with torch.cat(). Internally it uses torch.distributed.all_gather. The tricky part here is that it assumes that all processes create identically-sized tensors. If the sizes don't match, syncing will hang indefinitely.
Related
I replaced CIFAR-10 preprocessing pipeline in the project with Dataset API approach and it resulted in performance decrease of about 10-20%.
Preporcessing is rather standart:
- read image from disk
- make random/crop and flip
- shuffle, batch
- feed to the model
Overall i see that batche processing is now 15% faster, but every once in a while (or, more precisely, whenever I reinitialize dataframe or expect reshuffling) the batch is being blocked for up long time (30 sec) which totals to slower epoch-per-epoch processing.
This behaviour seems to do something with internal hashing. If I reduce N in ds.shuffle(buffer_size=N) delays are shorter but proportionally more frequent. Removing shuffle at all results to delays as if buffer_size was set to dataset size.
Can somebody explain internal logic of Dataset API when it comes to reading/caching? Is there any reason at all to expect Dataset API to work faster than manually created Queues?
I am using TF 1.3.
If you implement the same pipeline using the tf.data.Dataset API and using queues, the performance of the Dataset version should be better than the queue-based version.
However, there are a few performance best practices to observe in order to get the best performance. We have collected these in a performance guide for tf.data. Here are the main issues:
Prefetching is important: the queue-based pipelines prefetch by default and the Dataset pipelines do not. Adding dataset.prefetch(1) to the end of your pipeline will give you most of the benefit of prefetching, but you might need to tune this further.
The shuffle operator has a delay at the beginning, while it fills its buffer. The queue-based pipelines shuffle a concatenation of all epochs, which means that the buffer is only filled once. In a Dataset pipeline, this would be equivalent to dataset.repeat(NUM_EPOCHS).shuffle(N). By contrast, you can also write dataset.shuffle(N).repeat(NUM_EPOCHS), but this needs to restart the shuffling in each epoch. The latter approach is slightly preferable (and truer to the definition of SGD, for example), but the difference might not be noticeable if your dataset is large.
We are adding a fused version of shuffle-and-repeat that doesn't incur the delay, and a nightly build of TensorFlow will include the custom tf.contrib.data.shuffle_and_repeat() transformation that is equivalent to dataset.shuffle(N).repeat(NUM_EPOCHS) but doesn't suffer the delay at the start of each epoch.
Having said this, if you have a pipeline that is significantly slower when using tf.data than the queues, please file a GitHub issue with the details, and we'll take a look!
Suggested things didn't solve my problem back in the days, but I would like to add a couple of recommendations for those, who don't want to learn about queues and still get the most out of TF data pipeline:
Convert your input data into TFRecord (as cumbersome as it might be)
Use recommended input pipeline format
.
files = tf.data.Dataset.list_files(data_dir)
ds = tf.data.TFRecordDataset(files, num_parallel_reads=32)
ds = (ds.shuffle(10000)
.repeat(EPOCHS)
.map(parser_fn, num_parallel_calls=64)
.batch(batch_size)
)
dataset = dataset.prefetch(2)
Where you have to pay attention to 3 main components:
num_parallel_read=32 to parallelize disk IO operations
num_parallel_calls=64 to parallelize calls to parser function
prefetch(2)
I know that some of Spark Actions like collect() cause performance issues.
It has been quoted in documentation
To print all elements on the driver, one can use the collect() method to first bring the RDD to the driver node thus:rdd.collect().foreach(println). This can cause the driver to run out of memory, though,
because collect() fetches the entire RDD to a single machine; if you only need to print a few elements of the RDD, a safer approach is to use the take(): rdd.take(100).foreach(println).
And from one more related SE question: Spark runs out of memory when grouping by key
I have come to know that groupByKey(), reduceByKey() may cause out of memory if parallelism is not set properly.
I did not get enough evidence on other Transformations and Action commands, which have to be used with caution.
These three are the only commands to be tackled? I have doubts about below commands too
aggregateByKey()
sortByKey()
persist() / cache()
It would be great if you provide information on intensive commands (global across partitions instead of single partition OR low performance commands), which have to be tackled with better guarding.
You have to consider three types of operations:
transformations implemented using only mapPartitions(WithIndex) like filter, map, flatMap etc. Typically it will be the safest group. Probably the biggest possible issue you can encounter is an extensive spilling to disk.
transformations which require shuffle. It includes obvious suspects like different variants of combineByKey (groupByKey, reduceByKey, aggregateByKey) or join and less obvious like sortBy, distinct or repartition. Without a context (data distribution, exact function for reduction, partitioner, resources) it is hard to tell if particular transformation will be problematic. There are two main factors:
network traffic and disk IO - any operation which is not performed in memory will be at least an order of magnitude slower.
skewed data distribution - if distribution is highly skewed shuffle can fail or subsequent operations may suffer from a suboptimal resource allocation
operations which require passing data to and from the driver. Typically it covers actions like collect or take and creating distributed data structure from a local one (parallelize).
Other members of this category are broadcasts (including automatic broadcast joins) and accumulators. Total cost depends of course on a particular operation and the amount of data.
While some of these operations can be expensive none is particularly bad (including demonized groupByKey) by itself. Obviously it is better to avoid network traffic or additional disk IO but in practice you cannot avoid it in any complex application.
Regarding cache you may find Spark: Why do i have to explicitly tell what to cache? useful.
Please bear with me, this is a basic architectural question for my first attempt at a "big data" project, but I believe your answers will be of general interest to anyone who is starting out in this field.
I've googled and read the high-level descriptions of Kafka, Storm, Memcached, MongoDB, etc., but now that I'm ready to dig in to start designing my app, I still need some further insight on how in fact the data should be distributed and shared.
The performance of my app is critical, so one objective is to somehow maximize the locality of the data in the RAM of the machines doing the distributed calculations. I need advice for this part of the design.
If my app had some clear criteria for a priori sharding the data and distributing the calculations (such as geographical regions or company divisions) then the solution would be obvious. But unfortunately my app's data access patterns are dynamic and depend on the results of previous calculations.
My app is an analysis program with distinct stages. In the first stage, all the data is accessed once and a metric is calculated for each data object. In the second stage, a subset of the data objects may be accessed, with the probability of access being proportional to each data object's metric that was calculated in the previous stage. In the final stage, a relatively small subset of data objects will be accessed many times for many calculations.
At all stages, it is required that the calculations be distributed across several servers. The calculations are embarassingly parallel, and each distributed calculation only needs to access a few data objects. It is also required that the number of servers can be specified before the app runs (for example, run on one server, or run on fifty servers).
It seems to me that I need some mechanism that distributes the appropriate data objects to the appropriate compute servers, as opposed to just blindly fetching the data from some database service (whether centralized or distributed). Also, it seems to me that some sort of smart caching system might be appropriate, since the data access pattern depends on the previous calculations and cannot be predicted a priori. But as far as I can tell, Memcached is not such a system because the sharding is determined a priori.
I've read many times that the operating system cache performs better than any monkeying around that we may try. I think the ideal solution is that each compute server's RAM cache somehow captures the data objects' dynamic access patterns, but it's not clear to me how this would work with a NoSQL or Memcached service.
Thanks for bearing with me this far. I realize this is a basic question, but the answer eludes me so far. I can't resolve the dynamic access patterns of my app with the a priori sharding of the NoSQL/Memcached packages. Any advice would be greatly appreciated.
I recommend you to take a look at http://tarantool.org. Shard to maximize locality for the most common data access pattern, use Lua for local computations, and net.box to issue a remote RPC when calculation needs to continue on another node. All data is stored in RAM, if you write your computation code carefully it could take advantage of the Just In Time compiler.
My company currently services their clients using a Windows-based fat-client app that has workflow processing embedded in it. Basically, the customer inserts a set of documents to the beginning of the workflow, the documents are processed through a number of workflow steps and then after a period of time the output is presented to the customer. We currently scale up for larger customers by installing the application on other machines and let the cluster of machines work on different document subsets. Not ideal but with minimal changes to the application it did allow us to easily scale up to our current level.
The problem we now face is that as our customers have provided us with larger document sets we find ourselves spending more than expected on machines, IT support, etc... So, we have started to think about re-architecting the platform to make it scaleable. A feature of our solution is that each document can be processed independently of one another. Also we have 10 workflow steps of which two of the steps take up about 90% of the processing time.
One idea we are mulling over is to add a workflow step field to the document schema to track which workflow step has been completed for the document. We can then throw the entire cluster of machines to work on a single document set. A single machine would not be responsible for sequentially processing a document through all workflow steps but queries the db for the next document/workflow step pair and perform that processing. Does this sound like a reasonable approach? Any suggestions?
Thanks in advance.
While I'm not sure what specific development environment you are working with, I have had to deal with some similar workflows where we have a varied number of source documents, various steps, etc. all with different performance characteristics.
Assuming you have a series of independent steps - i.e. Step A's work product is the input for Step B, and step B's product is the input for step C, etc. I would look at message queueing as a potential solution.
For example, all new documents are tossed into a queue. One or more listener apps hit the queue and grab the next available document to perform step A. As step A completes, a link to the output product and/or relevant data are tossed into another queue. a separate listener app pulls from this second queue into step B, etc. until the final output product is created.
In this way, you use one queue for the holding area between each discreet step, and can scale up or down any individual process between the queues.
For example, we use this to go from some data transformations, through a rendering process, and out to a spooler. The data is fast, the renderings are CPU bound, and the printing is I/O bound, but each individual step can be scaled out based on need.
You could (technically) use a DB for this - but a message queue and/or a service bus would likely serve you better.
Hopefully that points you in the right direction!
Assuming I have a cluster of n Erlang nodes, some of which may be on my LAN, while others may be connected using a WAN (that is, via the Internet), what are suitable mechanisms to cater for a) different bandwidth availability/behavior (for example, latency induced) and b) nodes with differing computational power (or even memory constraints for that matter)?
In other words, how do I prioritize local nodes that have lots of computational power, over those that have a high latency and may be less powerful, or how would I ideally prioritize high performance remote nodes with high transmission latencies to specifically do those processes with a relatively huge computations/transmission (that is, completed work per message ,per time unit) ratio?
I am mostly thinking in terms of basically benchmarking each node in a cluster by sending them a benchmark process to run during initialization, so that the latencies involved in messasing can be calculated, as well as the overall computation speed (that is, using a node-specific timer to determine how fast a node terminates with any task).
Probably, something like that would have to be done repeatedly, on the one hand in order to get representative data (that is, averaging data) and on the other hand it might possibly even be useful at runtime in order to be able to dynamically adjust to changing runtime conditions.
(In the same sense, one would probably want to prioritize locally running nodes over those running on other machines)
This would be meant to hopefully optimize internal job dispatch so that specific nodes handle specific jobs.
We've done something similar to this, on our internal LAN/WAN only (WAN being for instance San Francisco to London). The problem boiled down to a combination of these factors:
The overhead in simply making a remote call over a local (internal) call
The network latency to the node (as a function of the request/result payload)
The performance of the remote node
The compute power needed to execute the function
Whether batching of calls provides any performance improvement if there was a shared "static" data set.
For 1. we assumed no overhead (it was negligible compared to the others)
For 2. we actively measured it using probe messages to measure round trip time, and we collated information from actual calls made
For 3. we measured it on the node and had them broadcast that information (this changed depending on the load current active on the node)
For 4 and 5. we worked it out empirically for the given batch
Then the caller solved to get the minimum solution for a batch of calls (in our case pricing a whole bunch of derivatives) and fired them off to the nodes in batches.
We got much better utilization of our calculation "grid" using this technique but it was quite a bit of effort. We had the added advantage that the grid was only used by this environment so we had a lot more control. Adding in an internet mix (variable latency) and other users of the grid (variable performance) would only increase the complexity with possible diminishing returns...
The problem you are talking about has been tackled in many different ways in the context of Grid computing (e.g, see Condor). To discuss this more thoroughly, I think some additional information is required (homogeneity of the problems to be solved, degree of control over the nodes [i.e. is there unexpected external load etc.?]).
Implementing an adaptive job dispatcher will usually require to also adjust the frequency with which you probe the available resources (otherwise the overhead due to probing could exceed the performance gains).
Ideally, you might be able to use benchmark tests to come up with an empirical (statistical) model that allows you to predict the computational hardness of a given problem (requires good domain knowledge and problem features that have a high impact on execution speed and are simple to extract), and another one to predict communication overhead. Using both in combination should make it possible to implement a simple dispatcher that bases its decisions on the predictive models and improves them by taking into account actual execution times as feedback/reward (e.g., via reinforcement learning).