Parallel Execution

This page covers PARALLEL FOR, the slice property system, and parallel reduction functions. They are the key building blocks for high-performance graph algorithms.

PARALLEL FOR

Execute loop iterations across multiple worker goroutines:

Basic Syntax

Procedure Body Language
-- Auto-detect worker count (uses all CPU cores)
PARALLEL FOR n IN SCAN(:Person) {
    n.processed = true
}

Explicit Worker Count

Procedure Body Language
PARALLEL FOR n IN SCAN(:Person) WORKERS 8 {
    LET score = OUT_DEGREE(n) * 0.1
    n.score = score
}

With Batching

Use .batch(N) for better throughput — nodes are fetched in batches of N:

Procedure Body Language
PARALLEL FOR n IN SCAN(:Person).batch(1000) WORKERS 4 {
    LET degree = OUT_DEGREE(n)
    SET_SLICE_PROP(n._internal_id, 'degree', degree)
}

Thread Safety Notes

• Slice properties (GET_SLICE_PROP, SET_SLICE_PROP) are safe for parallel access, each node has its own slot indexed by _internal_id.
• Temp properties (node.prop = value) are safe — assigned per-node.
• LET variables inside the loop body are local to each iteration.
• Shared accumulators (e.g., LET count = count + 1) may have race conditions. For reductions, use SUM_SLICE_PROP instead.
• RETURN statements are thread-safe, results are collected atomically.

Slice Property System

Slice properties provide O(1) per-node value storage backed by contiguous arrays indexed by _internal_id. They are the primary data structure for implementing graph algorithms.

_internal_id

Every node has an _internal_id, a system-assigned integer used as the array index for slice properties. Access it via node._internal_id:

Procedure Body Language
PARALLEL FOR n IN SCAN() WORKERS 8 {
    -- n._internal_id is the array index
    SET_SLICE_PROP(n._internal_id, 'rank', 1.0)
}

INIT_SLICE_PROP

Initialize a slice property with a uniform value for all nodes:

Procedure Body Language
-- Initialize all nodes with rank = 1/N
LET n = NODE_COUNT()
INIT_SLICE_PROP('rank', 1.0 / n)

-- Initialize all nodes with component = 0
INIT_SLICE_PROP('component', 0.0)

Internally parallelized for large graphs (>10K nodes).

GET_SLICE_PROP

Read a node's slice property value by internal ID:

Procedure Body Language
LET rank = GET_SLICE_PROP(node._internal_id, 'rank')

Also accepts integer index directly:

Procedure Body Language
LET val = GET_SLICE_PROP(42, 'score')

SET_SLICE_PROP

Write a node's slice property value:

Procedure Body Language
SET_SLICE_PROP(node._internal_id, 'rank', new_rank)

COPY_SLICE_PROP

Copy all values from one slice to another (bulk operation):

Procedure Body Language
-- Copy new_rank → rank (for iterative algorithms)
COPY_SLICE_PROP('new_rank', 'rank')

Internally parallelized for large slices. Useful between iterations of algorithms like PageRank.

INIT_OUT_DEGREES

Initialize a slice with out-degree values from the topology accelerator:

Procedure Body Language
INIT_OUT_DEGREES('out_degree')

-- Now GET_SLICE_PROP(node._internal_id, 'out_degree') returns the out-degree

This is much faster than computing degrees in a loop.

Parallel Reduction Functions

High-performance aggregate operations over slice properties. Automatically parallelized.

SUM_SLICE_PROP

Sum all values in a slice property:

Procedure Body Language
LET total_rank = SUM_SLICE_PROP('rank')

SUM_SLICE_PROP_SQ

Sum of squares - essential for L2 normalization:

Procedure Body Language
LET norm_sq = SUM_SLICE_PROP_SQ('score')
LET norm = SQRT(norm_sq)

MAX_SLICE_PROP / MIN_SLICE_PROP

Find extremes across all nodes:

Procedure Body Language
LET max_rank = MAX_SLICE_PROP('rank')
LET min_rank = MIN_SLICE_PROP('rank')
LET range = max_rank - min_rank

Batch Persistence

Persist slice property values to actual node properties in storage.

BATCH_PERSIST_SLICE

Persist a single slice to storage:

Procedure Body Language
-- After PageRank completes, save results
BATCH_PERSIST_SLICE('rank', 'pagerank_score')

This writes slice['rank'][node._internal_id] → node.pagerank_score for all nodes.

BATCH_PERSIST_SLICES

Persist multiple slices in a single efficient pass:

Procedure Body Language
-- HITS algorithm: persist both hub and authority scores
BATCH_PERSIST_SLICES('hub', 'hub_score', 'auth', 'authority_score')

BATCH_PERSIST_SLICES is more efficient than multiple BATCH_PERSIST_SLICE calls because it:

• Reads from property cache only once per node.
• Copies property map only once per node.
• Encodes and writes only once per node.

BATCH_SLICE_ADD_SCALAR

Add a constant value to every element in a slice in place:

Procedure Body Language
-- Add teleportation probability to all ranks
BATCH_SLICE_ADD_SCALAR('rank', (1.0 - damping) / n)

BATCH_SLICE_ADD_SCALAR(sliceName, constant): for each node i, slice[i] = slice[i] + constant.

BATCH_SLICE_ADD

Element-wise addition of two slices into a third slice:

Procedure Body Language
-- result[i] = slice1[i] + slice2[i]
BATCH_SLICE_ADD('slice1', 'slice2', 'result')

BATCH_SLICE_ADD(slice1Name, slice2Name, outputSliceName). The output slice must already be initialized (e.g., with INIT_SLICE_PROP).

Pattern: Iterative Algorithm

Most graph algorithms follow this pattern using the parallel execution primitives:

GQL
CREATE PROCEDURE algorithm(iterations: INT = 20)
RETURNS (node_id: STRING, score: FLOAT)
AS {
    -- 1. Initialize slice properties
    INIT_SLICE_PROP('score', initial_value)
    INIT_SLICE_PROP('new_score', 0.0)
    INIT_OUT_DEGREES('out_degree')

    -- 2. Iterate until convergence or max iterations
    FOR iter IN RANGE(0, $iterations) {
        -- 2a. Reset accumulator
        INIT_SLICE_PROP('new_score', 0.0)

        -- 2b. Parallel computation
        PARALLEL FOR n IN SCAN() WORKERS 8 {
            LET contrib = IN_NEIGHBOR_SUM(n, 'score', 'out_degree')
            SET_SLICE_PROP(n._internal_id, 'new_score', contrib)
        }

        -- 2c. Swap (or normalize then swap)
        COPY_SLICE_PROP('new_score', 'score')
    }

    -- 3. Persist results
    BATCH_PERSIST_SLICE('score', 'algorithm_score')

    -- 4. Return results
    FOR n IN SCAN() {
        LET score = GET_SLICE_PROP(n._internal_id, 'score')
        RETURN n._id AS node_id, score
    }
}

Performance Tips

Choose the Right Worker Count

• CPU-bound (arithmetic, slice ops): Use WORKERS N where N = number of CPU cores
• I/O-bound (property lookups from storage): Use more workers (2x-4x cores)
• Default (no WORKERS specified): Auto-detects CPU count

Use .batch() for Large Scans

Batching reduces scheduling overhead:

Procedure Body Language
-- Without batching: each node is dispatched individually
PARALLEL FOR n IN SCAN(:Person) WORKERS 8 { ... }

-- With batching: nodes dispatched in chunks of 1000
PARALLEL FOR n IN SCAN(:Person).batch(1000) WORKERS 8 { ... }

Prefer Slice Properties Over Temp Properties

For algorithms that access properties in PARALLEL FOR:

Procedure Body Language
-- FAST: O(1) array access
SET_SLICE_PROP(n._internal_id, 'rank', value)
LET rank = GET_SLICE_PROP(n._internal_id, 'rank')

-- SLOWER: Map-based property access
n.rank = value
LET rank = n.rank

Use Fused Neighbor Operations

Instead of manually iterating neighbors:

Procedure Body Language
-- SLOW: Per-neighbor interpreter overhead
LET sum = 0
FOR neighbor IN NEIGHBORS(node, IN) {
    LET sum = sum + GET_SLICE_PROP(neighbor._internal_id, 'rank')
}

-- FAST: Single fused operation, direct CSR/CSC access
LET sum = SUM_IN_NEIGHBOR_PROP(node, 'rank')