DATABASE OPERATORS AND SOLID STATE DRIVES Geetali Tyagi ( ) Mahima Malik ( ) Shrey Gupta ( ) Vedanshi Kataria ( )

Introduction ■ SSDs have rich internal parallelism – Chance to improve I/O bandwidth by doing parallel processing ■ Higher IOPS (Input/Output Operations Per Second) ■ Traditional query processing algorithms mainly designed according to the mechanical traits of the HDDs – Simply replacing HDDs by SSDs does very little benefit – Redesign algorithms like scan and join to take full advantage of SSDs

SSD Internal Architecture ■ Multiple channels shared by a set of flash memory packages ■ Two levels of parallelism – Channel level parallelism ■ Each channel can be operated independently and simultaneously – Package level parallelism ■ Operations on flash memory packages attached to the same channel can also be interleaved

Some Terms ■ Domain is a set of flash memories that share a specific set of resources (e.g. channels) – Can further be partitioned into sub-domains ■ Chunk is a logical data page with a unique logical address. ■ Every domain has into its own ScanBuffer to store scan results for the particular domain.

ParaScan ■ Most SSDs adopt a RAID-0 like striping data storage mechanism – with consecutive Logical Block Addresses for the striped chunks ■ Try uniform distribution of the table across domains to maximize parallelism benefits. ■ Example: On the assumption of 20 domains, the chunk whose logical address is 20*n (n=0,1,2...), will be in 1st domain, the chunk whose logical address is 20*n+1 (n=0,1,2...), will be in 2nd domain, and so on. ■ Parascan is twice as fast as a traditional table scan on SSD and 4 times as fast as a traditional table scan on HDD in best case.

Domain 0 Domain 1 Chunk 0 Chunk 1 Chunk 2 Chunk 3

ParaScan ■ Domain scan – Read data chunks one by one from a single domain and then put them into its own ScanBuffer, allowing multiple domain scans to be executed in parallel without interference ■ Multi-domain parallel scan – Multiple threads ■ Each in charge of one or more individual domain scans – Entire scan buffer is also divided into several ScanBuffers so that each scan thread can use one ScanBuffer ■ Performance of multi-domain parallel scan depends on the concurrency level – the maximal number of physical threads supported by the processor – the maximal queue depth supported by the SSD

Buffer Multi-Domain Parallel Scan SSD Domain #0 ParaScan Chunk # Chunk #20 Chunk #200 Domain #1 Chunk # Chunk #21 Chunk # Domain #19 Chunk # Chunk #39 Chunk #219 Domain Scan ScanBuffer #0 Page #0Page #1 ScanBuffer #1 Page #0Page #1 ………… Page #0Page #1 ScanBuffer #19 Page #0Page #1

ParaHashJoin - Parallel Hash Join ■ Two-way equi-join consists of three phases – ParaHash, MiniJoin, and Fetch phase. ■ 3x times faster than traditional hash join in SSD Table R Table S ParaScan ParaHash MiniJoinFetch ParaHashJoi n

ParaHash ▪ Buffer - scan area and hash area. ▪ Multiple hash threads - to calculate the hash values of records in ScanBuffers  Each thread is assigned to one ScanBuffer ▪ Based on hash values, put their join attributes and RIDs into corresponding hash buckets in parallel  Hash function : Hash-value = join_attr & (B-1) ▪ Concurrency control  Each bucket maintain lightweight clock  Bitmap is used to check whether hash index records with specified join attribute exists in the bucket

ParaHashJoin ParaHash B-2B-1... Table Scan Area Hash Area

ParaHashJoin MiniJoin ▪ Input - ParaHash table R and ParaHash table S ▪ Each bucket is read into the memory to generate join results - {join_attr, RID R, RID S }  Two passes are required to generate MiniJoin results – one pass for ParaScan and one pass for MIniJoin ▪ If enough memory present to hold hash table of R (smaller table)  Table R is ParaScan and then ParaHash, Table S needs to be ParaScan only and can directly probe hash table of R for join results  Only one pass is required

ParaHashJoin Fetch ▪ Outputs necessary attributes using RIDs specified in the MiniJoin output index to get the final join results. ▪ TID Hash Join approach  For each join result, fetch the needed data pages to generate the final join result ▪ this approach is reasonable if all the pages of the result can fit in memory  When memory is insufficient, some pages can be loaded multiple times, resulting in higher cost of loading pages ▪ Sort-based fetching approach  Sort MiniJoin results based on the RIDs of outer table  Load needed pages to produce final join results according to the sorted MiniJoin results

ParaAggr (Parallel Aggregation) ■ Parallel implementation of aggregation operations (sum, max, min, count, average). ■ Two Phases: – SubAggr: Multiple threads corresponding to each ParaScan thread or ScanBuffer. – TolAggr: Combines results of all SubAggr instances. ■ ParaAggr in SSDs is 3-4 times faster than traditional aggregation using single thread in HDD.

ParaAggr – Working Diagram ParaAggr Domain # Domain #n ParaScan ParaScan SubAggr TotAggr SubAggr

ParaAggr - Example ParaAggr Domain # Domain #n ParaScan ParaScan SubAggr TotAggr SubAggr Scan records in all domains in parallel. Forward those satisfying WHERE clause to SubAggr. SELECT count(*) FROM Employee WHERE dept=’Sales’

ParaAggr - Example Count of ParaScan results in parallel SubAggr threads. SELECT count(*) FROM Employee WHERE dept=’Sales’ ParaAggr Domain # Domain #n ParaScan ParaScan SubAggr TotAggr SubAggr

ParaAggr - Example Summation on result of count SubAggr operations. SELECT count(*) FROM Employee WHERE dept=’Sales’ ParaAggr Domain # Domain #n ParaScan ParaScan SubAggr TotAggr SubAggr

Summary ■ Rich internal parallelism in SSDs exploited for faster data query processing. ■ Algorithms use parallel threads and SSD domains to speed up process. – E.g. ParaScan, ParaHashJoin, ParaSort, ParaAggr. ■ Para-SSD algorithms much faster that traditional HDD algorithms.

References ■ Y. Fan, W. Lai, X. Meng, Optimizing Database Operators by Exploiting Internal Parallelism of Solid State Drives