## What changes were proposed in this pull request?
Currently the whole block is fetched into memory(off heap by default) when shuffle-read. A block is defined by (shuffleId, mapId, reduceId). Thus it can be large when skew situations. If OOM happens during shuffle read, job will be killed and users will be notified to "Consider boosting spark.yarn.executor.memoryOverhead". Adjusting parameter and allocating more memory can resolve the OOM. However the approach is not perfectly suitable for production environment, especially for data warehouse.
Using Spark SQL as data engine in warehouse, users hope to have a unified parameter(e.g. memory) but less resource wasted(resource is allocated but not used). The hope is strong especially when migrating data engine to Spark from another one(e.g. Hive). Tuning the parameter for thousands of SQLs one by one is very time consuming.
It's not always easy to predict skew situations, when happen, it make sense to fetch remote blocks to disk for shuffle-read, rather than kill the job because of OOM.
In this pr, I propose to fetch big blocks to disk(which is also mentioned in SPARK-3019):
1. Track average size and also the outliers(which are larger than 2*avgSize) in MapStatus;
2. Request memory from `MemoryManager` before fetch blocks and release the memory to `MemoryManager` when `ManagedBuffer` is released.
3. Fetch remote blocks to disk when failing acquiring memory from `MemoryManager`, otherwise fetch to memory.
This is an improvement for memory control when shuffle blocks and help to avoid OOM in scenarios like below:
1. Single huge block;
2. Sizes of many blocks are underestimated in `MapStatus` and the actual footprint of blocks is much larger than the estimated.
## How was this patch tested?
Added unit test in `MapStatusSuite` and `ShuffleBlockFetcherIteratorSuite`.
Author: jinxing <jinxing6042@126.com>
Closes#16989 from jinxing64/SPARK-19659.
3f94e64aa8