A new runtime for k and q: l

Comments

[ 01A ] compatible k4, q, qSQL. Unchanged. Run the syntax, idioms, and code you already ship. Tables, dicts, and time-series work unchanged.

• k4 syntax compatibility
• Native q and qSQL support
• Optimized columnar tables
• Zero code rewrites

A new runtime for the k family

l runs k4, q, and qSQL unchanged - on a new engine built for modern memory: compressed vectors, SIMD, and automatic parallelism. The q/k/qSQL you already have, just faster.

l> sum 1 2 3 4 5 6
21

Modern HPC runtime: 0.02 ms SIMD reduction · scalar syntax

What is l

k and q made the vector the unit of thought. l keeps that language and makes the compressed vector the unit of execution: primitives run directly on the encoded bytes, so the full array is never rebuilt. Modern hardware starves for bandwidth long before it runs short of compute - so moving fewer bytes is what makes it fast.

[ 01A ] compatible - run the syntax, idioms, and code you already ship. Tables, dicts, and time-series work unchanged.

[ 01B ] compressed data stays compressed at rest, in memory, and in transit. Primitives run directly on the compressed vectors, so less data moves.

[ 01C ] transparent - the runtime picks the execution path - scalar, SIMD, threaded, or offloaded - with no annotations. You write sum x; l does the rest.

Performance

3.57x TSBS, master-benchmark, and db-benchmark results link back to their source repositories, result files, and validation notes. See All Benchmarks:

• 2.35x TSBS geomean vs reference
• 3.57x master suite overall geomean
• 3.16x DBB total query speedup
• 3.08x DBB per-query geomean

In practice

The same qSQL you write in q. Columns stay compressed while the engine runs over them.

// 86.5M-row analytics table - single CPU
l> count AN_pageView
86585107

l> select sum isBot, count i, avg isBot by date.month from AN_pageView
month    isBot   cnt      botRate
2024.05  29404   1634214  0.01799275
2024.06  49596   1495228  0.03316952
2024.07  194442  2812446  0.06913626

Single CPU · no decode pass · 86.5M rows

The same primitives, faster

One expression runs as a SIMD reduction on small vectors and fans out across threads on large ones.

l> avg 1 2 3 4 5 6
3.5
// -> fsum (SIMD reduction) -> / count

l> avg 1 2 3 4 ... 999999
// worker0: sum x[0..hi0]
// worker1: sum x[hi0..hi1]
// worker2: sum x[hi1..hi2]
// ...
// sum = s0 + s1 + s2 + ...
// avg = sum / n
499999.5

l> remote
Linux:hopen `:10.0.0.1:8080
connected

SIMD reduction · thread fan-out · remote execution · one expression

Anonymous mail group

In the spirit of classic technical mailing lists: notes, benchmarks, design questions, release logs. No profiles, no directories. Join with email only.