KFoxder / Alenka

GPU database engine

Welcome to Alenka - GPU database engine

How to build?

Download Alenka

Download CUDPP library

Download ModernGpu library

Install CUDPP and ModernGPU into Alenka directory

Run Makefile or Makefile.win

What is this?

This is a GPU based database engine written to use vector based processing and high bandwidth of modern GPUs

Features :

• Vector-based processing
  CUDA programming model allows a single operation to be applied to an entire set of data at once.

• Smart compression
  Ultra fast compression and decompression performed directly inside GPU.

• Column-based storage
  Minimize disk I/O by only accessing the relevant data.

• Data skipping
  Better performance without indexes.

How to use it ?

Create your data files :

Run scripts load_orders.sql, load_lineitem.sql and load_customer.sql to create your database files.

Step 1 - Load the data files

O := LOAD 'orders' BINARY AS (o_orderkey{1}:int, o_custkey{2}:int, o_orderdate{5}:int, o_shippriority{8}:int);

C := LOAD 'customer' BINARY AS (c_custkey{1}:int, c_mktsegment{7}:varchar(10));

L := LOAD 'lineitem' BINARY AS (l_orderkey{1}:int, price{6}:decimal, discount{7}:decimal, shipdate{11}:int);

Step 2 - Filter data

OFI := FILTER O BY o_orderdate < 19950315;

CF := FILTER C BY c_mktsegment == "BUILDING";

LF := FILTER L BY shipdate > 19950315;

Step 3 - Join data

OLC := SELECT o_orderkey AS o_orderkey, o_orderdate AS o_orderdate, o_shippriority AS o_shippriority, price AS price, discount AS discount FROM LF JOIN OFI ON l_orderkey = o_orderkey JOIN CF ON o_custkey = c_custkey;

Step 4 - Group data

F := SELECT o_orderkey AS o_orderkey1, o_orderdate AS orderdate1, o_shippriority AS priority, SUM(price*(1-discount)) AS sum_revenue, COUNT(o_orderkey) AS cnt
FROM OLC GROUP BY o_orderkey, o_orderdate, o_shippriority;

Step 5 - Order data

RES := ORDER F BY sum_revenue DESC, orderdate1 ASC;

Step 6 - Save the results

STORE RES INTO 'results.txt' USING ('|') LIMIT 10;

How fast is it?

Using TPC-H benchmark at the scale 1000 (1TB of data) we compared the following systems :

My system (Total cost : 1,700 USD) :

• CPU - Pentium G620, 2 cores
• GPU - NVidia GTX Titan, 6GB of DDR5 GPU memory
• 16GB of memory
• 1 Vertex3 120GB SSD
• AlenkaDB

Top #7 TPC-H result at scale 1000 (Total cost : $1,128,288 USD):

• IBM Power 780 Model 9179-MHB
• 8 IBM POWER7 4.1GHz CPUs, 32 cores
• 512 GB of memory
• 52 x 69GB SAS SSD
• Sybase IQ Single Application Server Edition v.15.2

Results of TPC-H Queries Q1,Q2 and Q3 in seconds ( the lower values are better )

	Q1	Q2	Q3
IBM+Sybase	118	7.8	27
Titan+Alenka	72	6.1	23

*Alenka run time calculated as total run time - disk time.

---