Hi Sunyu, Nice work! We've been working with your patch and enhancing it for incorporation into Apache Drill. What do you think the timeline and steps are to get this into master? We'd be more than happy to help depending on your time for this in the coming weeks.
thanks, Jacques On Thu, Aug 14, 2014 at 8:30 AM, sunyu duan <[email protected]> wrote: > Hi everyone, > > My name is Sunyu Duan, this year GSOC Student who working on Parquet. > > As most of the work has been done, I wrote this report to summarize what I > have done and the result. > > My Project is Using Zero-Copy read path in new Hadoop API. The goal is to > exploit the Zero-Copy API introduced by Hadoop to improve read performance > of parquet tasks running locally. My contribution is to replace byte array > based API with ByteBuffer based API in the reading path to avoid byte array > copy and keep compatible with old APIs. Here is the complete pull request. > https://github.com/apache/incubator-parquet-mr/pull/6 > > My work includes two parts. > > 1. Make the whole read path use ByteBuffer directly. > > > - Introduce an initFromPage interface in ValueRead and implement it in > each ValueReader. > - Introduce a ByteBufferInputStream. > - Introduce a ByteBufferBytesInput. > - Replace unpack8values method with a ByteBuffer version. > - Use introduced ByteBuffer based method in the read path. > > > 1. Introduce a Compatible layer to keep compatible with old Hadoop API > > > - Introduce a CompatibilityUtil > - Using the CompatiblityUtil to perform read action > > > > After coding, I started to benchmark the improvement. After discussion with > my mentor, I modified the TestInputOutputFormat test to inherit > ClusterMapReduceTestCase which will start a MiniCluster for unit test. In > the unit test, I enabled caching and read shortcircuiting. I created a > 500MB and a 1GB log file on my dev box for the test. The test will read in > the log file and write to the temporary parquet format file using > MapReduce. Then it will read from the temporary parquet format file and > write to an output file. I inserted time counter on the latter mapreduce > task and used the time spent on the seconde MapReduce Job as an indicator. > I ran the unit test with and without Zero-Copy API enabled on 500MB and 1GB > log file and compared the time spent on each situation. The result shows > below. > > > > File > Size Average Reading Time(s) Improvement > > Without Zero-Copy API 500MB > 576s > > Zero-Copy API > 500MB 394s > 46% > > Without Zero-Copy API 1024MB > 1080s > > Zero-Copy API 1024MB > 781s 38% > > > > As we can see, there is about 30~50% improvement on reading performance > which shows the project has reached its goal. But the benchmark is > insufficient. My dev box has very limited resources and 1GB file is the > maximum file I can put. After GSOC, it'd be better to invite more people to > try it out on real cluster with larger file to benchmark its effect on real > situation. > > > Best, > > Sunyu >
