Reynold, >From our experiments, it is not a massive refactoring of the code. Most expressions can be supported by a relatively small change while leaving the existing code path untouched. We didn't try to do columnar with code generation, but I suspect it would be similar, although the code generation code itself is rather complicated which is why we avoided it initially.
To address your other concerns. > 1. GPU machines are much more expensive & difficult to get. In our experiments, a 4x performance improvement in a query offsets the cost of a GPU, at least for the latest generation T4 GPUs. In some queries, we were able to do the entire processing for half the cost because we could finish faster. But yes it does depend on the type of processing you are doing and the more numerically complicated the processing is the more of a speedup you will see. One of the reasons that we started down this path is because our customers were seeing real-world ML use cases where the data preparation (ETL) was dominating the training time for their models. They already had GPU nodes to train their models on and were using Spark to do the data preparation and orchestrate the training so while ETL was happening they were paying for GPUs that were sitting idle, which was most of the time. Hopefully, the new GPU aware scheduling feature will remove the need for idle GPUs, but I still think the cost story is compelling even without it. And if the cost model makes sense, GPU enabled nodes will become easier to get ahold of. > 2. Bandwidth from system to GPUs is usually small This does vary a lot depending on the hardware you have. The theoretical limit of a 16 pair PCIe 3.0 connection is about 15 GB/s. In GCP we measured a 12GB/s bandwidth between GPU and main memory (bidirectional 12 up and 12 down simultaneously). Yes, this is tiny in comparison to the bandwidth you can get from L1 cache, but if you do need to hit main memory, which I would expect for big data, the bandwidth typically is not that different. PCIe 4.0 and 5.0 are coming and should surpass the theoretical limits of what DDR4 can deliver. So even if it does not make sense right now the cost calculations are likely to change over the next few years. You still take a hit because the data is typically staged in main memory before moving it to the GPU, which is why you need a cost-based optimizer to decide if it is worth it. We also have options to use compression as part of the transfers if necessary to increase the practical bandwidth. > 3. It's a massive effort. > In general it's a cost-benefit trade-off. If we include testing, documentation etc. yes it does add up to a lot of work. I get that you don't want someone to show up, dump a ton of untested code, and walk away. I don't want that either. That is why the proposal makes it a separate optional code path that can be off by default and allows us to add functionality one operator at a time. We are willing to make a long term commitment to this and see it through so that it is a success. If that means providing hardware for testing we already started discussions internally about what is the best way to do that. If it means ongoing support in the community we are willing to do that. Please let me know if I missed any of your concerns, or if I misunderstood any of them. Thanks, Bobby On Tue, Mar 26, 2019 at 12:21 PM Reynold Xin <r...@databricks.com> wrote: > 26% improvement is underwhelming if it requires massive refactoring of the > codebase. Also you can't just add the benefits up this way, because: > > - Both vectorization and codegen reduces the overhead in virtual function > calls > - Vectorization code is more friendly to compilers / CPUs, but requires > materializing a lot of data in memory (or cache) > - Codegen reduces the amount of data that flows through memory, but for > complex queries the generated code might not be very compiler / CPU friendly > > > I see massive benefits in leveraging GPUs (and other accelerators) for > numeric workloads (e.g. machine learning), so I think it makes a lot of > sense to be able to get data out of Spark quickly into UDFs for such > workloads. > > I don't see as much benefits for general data processing, for a few > reasons: > > 1. GPU machines are much more expensive & difficult to get (e.g. in the > cloud they are 3 to 5x more expensive, with limited availability based on > my experience), so it is difficult to build a farm > 2. Bandwidth from system to GPUs is usually small, so if you could fit the > working set in GPU memory and repeatedly work on it (e.g. machine > learning), it's great, but otherwise it's not great. > 3. It's a massive effort. > > In general it's a cost-benefit trade-off. I'm not aware of any general > framework that allows us to write code once and have it work against both > GPUs and CPUs reliably. If such framework exists, it will change the > equation. > > > > > > > On Tue, Mar 26, 2019 at 6:57 AM, Bobby Evans <bo...@apache.org> wrote: > >> Cloudera reports a 26% improvement in hive query runtimes by enabling >> vectorization. I would expect to see similar improvements but at the cost >> of keeping more data in memory. But remember this also enables a number of >> different hardware acceleration techniques. If the data format is arrow >> compatible and off-heap someone could offload the processing to native code >> which typically results in a 2x improvement over java (and the cost of a >> JNI call would be amortized over processing an entire batch at once). >> Also, we plan on adding in GPU acceleration and ideally making it a >> standard part of Spark. In our initial prototype, we saw queries which we >> could make fully columnar/GPU enabled being 5-6x faster. But that really >> was just a proof of concept and we expect to be able to do quite a bit >> better when we are completely done. Many commercial GPU enabled SQL >> engines claim to be 20x to 200x faster than Spark, depending on the use >> case. Digging deeply you see that they are not apples to apples >> comparisons, i.e. reading from cached GPU memory and having spark read from >> a file, or using parquet as input but asking spark to read CSV. That being >> said I would expect that we can achieve something close to the 20x range >> for most queries and possibly more if they are computationally intensive. >> >> Also as a side note, we initially thought that the conversion would not >> be too expensive and that we could just move computationally intensive >> processing onto the GPU piecemeal with conversions on both ends. In >> practice, we found that the cost of conversion quickly starts to dominate >> the queries we were testing. >> >> On Mon, Mar 25, 2019 at 11:53 PM Wenchen Fan <cloud0...@gmail.com> wrote: >> >> Do you have some initial perf numbers? It seems fine to me to remain >> row-based inside Spark with whole-stage-codegen, and convert rows to >> columnar batches when communicating with external systems. >> >> On Mon, Mar 25, 2019 at 1:05 PM Bobby Evans <bo...@apache.org> wrote: >> >> This thread is to discuss adding in support for data frame processing >> using an in-memory columnar format compatible with Apache Arrow. My main >> goal in this is to lay the groundwork so we can add in support for GPU >> accelerated processing of data frames, but this feature has a number of >> other benefits. Spark currently supports Apache Arrow formatted data as an >> option to exchange data with python for pandas UDF processing. There has >> also been discussion around extending this to allow for exchanging data >> with other tools like pytorch, tensorflow, xgboost,... If Spark supports >> processing on Arrow compatible data it could eliminate the >> serialization/deserialization overhead when going between these systems. >> It also would allow for doing optimizations on a CPU with SIMD instructions >> similar to what Hive currently supports. Accelerated processing using a GPU >> is something that we will start a separate discussion thread on, but I >> wanted to set the context a bit. >> >> Jason Lowe, Tom Graves, and I created a prototype over the past few >> months to try and understand how to make this work. What we are proposing >> is based off of lessons learned when building this prototype, but we really >> wanted to get feedback early on from the community. We will file a SPIP >> once we can get agreement that this is a good direction to go in. >> >> The current support for columnar processing lets a Parquet or Orc file >> format return a ColumnarBatch inside an RDD[InternalRow] using Scala’s type >> erasure. The code generation is aware that the RDD actually holds >> ColumnarBatchs and generates code to loop through the data in each batch as >> InternalRows. >> >> Instead, we propose a new set of APIs to work on an >> RDD[InternalColumnarBatch] instead of abusing type erasure. With this we >> propose adding in a Rule similar to how WholeStageCodeGen currently works. >> Each part of the physical SparkPlan would expose columnar support through a >> combination of traits and method calls. The rule would then decide when >> columnar processing would start and when it would end. Switching between >> columnar and row based processing is not free, so the rule would make a >> decision based off of an estimate of the cost to do the transformation and >> the estimated speedup in processing time. >> >> This should allow us to disable columnar support by simply disabling the >> rule that modifies the physical SparkPlan. It should be minimal risk to >> the existing row-based code path, as that code should not be touched, and >> in many cases could be reused to implement the columnar version. This also >> allows for small easily manageable patches. No huge patches that no one >> wants to review. >> >> As far as the memory layout is concerned OnHeapColumnVector and >> OffHeapColumnVector are already really close to being Apache Arrow >> compatible so shifting them over would be a relatively simple change. >> Alternatively we could add in a new implementation that is Arrow compatible >> if there are reasons to keep the old ones. >> >> Again this is just to get the discussion started, any feedback is >> welcome, and we will file a SPIP on it once we feel like the major changes >> we are proposing are acceptable. >> >> Thanks, >> >> Bobby Evans >> >> >
