Speed Q: Storing Large Graphs in DB - Inserting Several Million Rows - Research Proj.

[aeblank]

THE PROBLEM

I'm running into performance issues generating and storing a randomly created graph in a SQL Server database. I have a T-SQL script that generates a graph, and then randomly connects the vertices in that graph to each other. I can see that my hard-drive is working very hard throughout the operation. The operation took about 2 hours (I just canceled it at this point, looked like it was about 10% done) with 200,000 vertices and an average of 50 edges per vertex (essentially, it had to insert about 12 million rows).

How can I speed this up? Is there a clever way to get SQL Server to do all the operations in RAM, and then just copy the data from RAM to disk in one contiguous stream after all the inserting is done?

I am developing on a 4gb RAM laptop, and the database should easily fit entirely into RAM several times over--I estimate that my "test" script would generate a 150mbyte database.

THE BACKGROUND

I'm new to database programming and am trying to store a directed graph in a database for a research project. The graph will ultimately contain approximately 1,000,000 vertices with an average of about 150 edges per vertex. There will be more than one graph stored. When this is deployed, it will be deployed on a cluster with (much) more RAM, storage and computing power.

Each edge represents a uni-directional connection between two points (vertices) on the graph). So if point 10 connects to point 23 in a graph, there would be an "edge" entry in the database containing "10" and "23".

If a vertex/point is isolated (it connects to nothing), it is stored as an edge which "connects" to itself: i.e. if point 56 connects to no other points, then there would be an "edge" entry in the database containing "56" and "56".

THE SCHEMA

The table is of the form:

CREATE TABLE [dbo].[edges](
    [edge_id] [int] IDENTITY(1,1) NOT NULL,
    [graph_id] [int] NOT NULL,
    [primary_vertex_id] [int] NOT NULL,
    [adjacent_vertex_id] [int] NULL CONSTRAINT [DF_edges_adjacent_vertex_id]  DEFAULT (NULL),
 CONSTRAINT [PK_edges] PRIMARY KEY CLUSTERED 
(
    [edge_id] ASC
)WITH (PAD_INDEX  = OFF, IGNORE_DUP_KEY = OFF) ON [PRIMARY]
) ON [PRIMARY]
 

THE SCRIPT THAT TAKES TOO LONG

This was cobbled together from two functions ("add vertices" and "connect vertices"), I hope it isn't too difficult to read.

DECLARE @graph_id INT
DECLARE @vertices INT
DECLARE @average_edges INT
SET @graph_id = 35
SET @vertices = 200000 /* 200,000 vertex graph */
SET @average_edges = 50; /* approx 50 connections per vertex */
    BEGIN
        /* get the first new vertex id -- should usually be 1, but just in case there's already some vertices in graph */
        DECLARE @first_vertex_id INT
        SELECT @first_vertex_id = (1 + isnull(max(dbo.edges.primary_vertex_id),0)) FROM dbo.edges WHERE dbo.edges.graph_id = @graph_id;
        BEGIN
            DECLARE @counter INT
            DECLARE @new_vertex_id INT
            SET @counter = 0
            WHILE @counter < @vertices
            BEGIN
                SET @new_vertex_id = @first_vertex_id + @counter;
                SET @counter = @counter + 1;
                /* insert a partial edge with the new vertex id */
                INSERT INTO dbo.edges(graph_id,
                    primary_vertex_id,
                    adjacent_vertex_id)
                VALUES(@graph_id, @new_vertex_id, @new_vertex_id);
            END;
            RETURN 0;
        END;
    END;
 
    BEGIN
        DECLARE @lower INT
        DECLARE @upper INT
        SELECT @lower = min(dbo.edges.primary_vertex_id) FROM dbo.edges WHERE dbo.edges.graph_id = @graph_id
        SELECT @upper = max(dbo.edges.primary_vertex_id) FROM dbo.edges WHERE dbo.edges.graph_id = @graph_id
 
        /* note: this algorithm assumes that the vertex_ids are CONTIGUOUS!! and may create duplicate edges */
        BEGIN
            DECLARE @loop_max INT
            DECLARE @counter INT
            DECLARE @rn_one INT
            DECLARE @rn_two INT
            SET @counter = 0
            SELECT @loop_max = (@upper - @lower + 1) * @average_edges
            WHILE @counter < @loop_max
            BEGIN
                /* pick two random vertex IDs that exist on the graph */
                SELECT @rn_one = Round(((@Upper - @Lower -1) * Rand() + @Lower), 0);
                SELECT @rn_two = Round(((@Upper - @Lower -1) * Rand() + @Lower), 0);                
                /* insert a bi-directional edge for those two vertices */
                INSERT INTO dbo.edges(graph_id,
                    primary_vertex_id,
                    adjacent_vertex_id)
                VALUES(@graph_id, @rn_one, @rn_two);
 
                INSERT INTO dbo.edges(graph_id,
                    primary_vertex_id,
                    adjacent_vertex_id)
                VALUES(@graph_id, @rn_two, @rn_one);
                SET @counter = @counter + 1;
            END
        END;
    END;
 

[Delerna]

Is there a clever way to get SQL Server to do all the operations in RAM, and then just copy the data from RAM to disk in one contiguous stream after all the inserting is done

declare @RamTable table (Field1 int,Field2 varchar(50))
--You now have a table that exists in ram
 
insert into @RamTable select 1,'string value'
insert into @RamTable  1,'string value'
update @RamTable  set Field2=Field2 + convert(varchar(3),Field1)
--Some silly sql to show that you use it like any real table
 
insert into TheRealTable
select * from @RamTable  
--Now you can save the results of all your processing into the actual table
 

Whether this works any faster or not I will leave for you to discover.
Hope it helps?

[aeblank]

Your tip works great: this sped it up by MORE THAN AN ORDER OF MAGNITUDE ! Thank you very much!

Details:
A 100,000 vertex graph, with an average of 100 connections per vertex, executed fully in 10:52 (just under 11 minutes!), although it used about 1035MB of RAM. The total # of edges in the DB is now approximately 20 million.