External Sorting algorithm for Databases having constrained storage hierarchy

• Divy Patel (9085310937) dspatel6@wisc.edu
• Sahil Naphade (9085746619) smnaphade@wisc.edu
• Devaki Kulkarni (9086222321) dgkulkarni2@wisc.edu
• Manaswini Gogineni (9085432699) mgogineni@wisc.edu

• Divy: Cache-size mini runs, Device-optimized page sizes, Spilling memory-to-SSD, Spilling from SSD to disk, Graceful degradation, Optimized merge patterns, Testing and Memory Leak Check
• Sahil: Tournament trees, Offset-value coding, Minimum count of row & column comparisons, Optimized merge patterns, Large-size records, Testing and Memory Leak Check
• Devaki: Tournament trees, Offset-value coding, Large-size records
• Manaswini: Verification

• Tournament trees: File Tree.cpp @ Line 196

• Offset-value coding: File DataRecord.cpp @ Line 122

• Minimum count of row & column comparisons

• Cache-size mini runs: File SortRecords.cpp @ Line 26

• Device-optimized page sizes: File SortRecords.cpp @ Line 81 and Line 136

• Spilling memory-to-SSD: File SortRecords.cpp @ Line 65

• Spilling from SSD to disk: File SortRecords.cpp @ Line 69 and Line 125

• Graceful degradation: File SortRecords.cpp @ Line 72, Line 74 and Line 151

  • Into merging
  • Beyond one merge step

• Optimized merge patterns: File SortRecords.cpp @ Line 150 and Line 151

• Verifying: File Iterator.cpp @ Line 69 and Line 84

  • sets of rows & values: File Iterator.cpp @ Line 84
  • sort order: File Iterator.cpp @ Line 69

• Replacement selection?

• Run size > memory size?

• Variable-size records?

• Compression?

• Prefix truncation?

• Quicksort

• Tournament-tree priority queue was used in order to achieve high fan-in for merging our sorted run inputs of records and less number of comparisons than a standard tree-of-winners
• Offset-value coding was used to achieve minimum column value comparisons
• Cache-size mini runs were used to be able to fit the sort inputs, for tournament-tree, in the cache. This enabled us to leverage the low-latency accesses when there are cache hits
• Device-optimized page sizes were used in order to being cognizant about the access-profile(latency, bandwidth) of various devices in the storage hierarchy. For SSD, we used 8KB(100 MB/s * 0.1 ms ~ 10KB) and for HDD, we used 1MB(100 MB/s * 10 ms ~ 1MB)
• We achieved graceful-degradation by spilling cache-size runs from cache to memory, spilling memory-size runs from memory to SSD and spilling SSD-size runs from SSD to HDD
• Also HDD-page size(1MB) sorted runs were written to SSD prior to actually merging runs on the HDD. This is to leverage low-latency accesses of flash drives(SSD)
• Sort-order, set of rows and their values were verified as part of sorting the input records. This is to verify the correctness and integrity of our sort algorihthm

• The implementation of the External-Sort is complete with all of the techniques which were expected from us as part of the course project
• The sort was tested against 1KB size records and with 12M number of records(although it takes ~1hr to complete the sort, for this particular test-case)
• The sort algorithm was tested against valgrind to check for any memory leaks introduced while developing. The codebase does not have any memory leaks, from the latest leak-report on the most recent code version

• To run our program, first compile the source code using following command, under External-Sort directory

$ cd External-Sort $ make ExternalSort.exe 

• After compiling the source code, to execute the External Sort with custom arguments, run following command inside External-Sort directory

# Where, # "-c" gives the total number of records # "-s" is the individual record size # "-o" is the trace of your program run $ ./ExternalSort.exe -c 120 -s 10 -o trace0.txt 

• The program creates three directories on the completion of the sort algorithm:

  • input: This directory consist of the input table which has records generated by the random-generator in arbitrary order
  • output: This directory consist of the output table which has records from input table but in a sorted order, sorted using our sort algorithm
  • trace: This directory consists of trace files generated from the sort. The trace file consists of logs related SSD and HDD device accesses. And the logs related to sort state machine

• In order to remove all the generated binaries, executables, and the utility directories mentioned above, run the following command

$ make clean 

$ docker run -it --privileged -v $pwd/External-Sort:/External-Sort ubuntu bash $ apt-get update $ apt-get install build-essential make g++ vim sudo valgrind -y $ cd ./External-Sort 

$ make ExternalSort.exe # Where, # "-c" gives the total number of records # "-s" is the individual record size # "-o" is the trace of your program run $ ./ExternalSort.exe -c 120 -s 1000 -o trace0.txt 

# Creates a log file `valgrind` inside the External-Sort directory $ valgrind --track-origins=yes --log-file="/External-Sort/valgrind" --leak-check=yes ./ExternalSort.exe -c 120 -s 1000 -o trace0.txt