High-throughput transport benchmark

Profiling session across concurrency levels, memory bounds, and latency plateaus.

Executive benchmark summary

The connection-pooled transport operates at 18x lower CPU utilization and maintains resident memory within 120 MB under peak load, compared to 816 MB in the baseline single-threaded runner.

Scaling worker concurrency from 1 to 4 sessions reduces overall sweep duration from 151 s down to 35 s. Concurrency beyond 8 worker sessions encounters host round-trip serialization and yields diminishing returns. Four worker threads is the recommended production setting.

Workload throughput and resource utilization

Table 1. Resource profiling comparing baseline vs optimized pool configurations across 300 test units.
Workload scenario	Baseline	Direct serial	Optimized pool (2)	Optimized pool (4)	Optimized pool (8)
Sweep duration (wall s)	2,772	151	78	35	32
Processor time (CPU s)	169.0	15.5	7.7	6.9	6.8
Peak resident set (RSS MB)	816	120	111	126	142
Range scan duration (wall s)	640	228	110	50	39
Range peak memory (MB)	1,436	297	245	223	289

Resource telemetry (RSS over elapsed time, CPU over sweep progress)

baseline-single-thread (peak 816 MB)

baseline-single-thread

pool-2-sessions (peak 111 MB)

pool-2-sessions

pool-4-sessions (peak 126 MB)

pool-4-sessions

pool-8-sessions (peak 142 MB)

pool-8-sessions

pool-2, pool-4, pool-8 (mean 9.9, 19.7 and 21.3 %)

pool-2 pool-4 pool-8

Key architectural optimizations

Iterative tree traversal. Replaced recursive generator tree traversals with an iterative loop and revision caching, eliminating 330 million generator frames per 200 requests.
Cyclic reference cleanup. Explicitly clearing parent and child document links on node replacement allows the garbage collector to immediately release 22 MB buffers by reference counting.
Shared authentication session pool. Pre-authenticates worker threads in parallel and dispatches chunked workload items from a thread-safe synchronized queue.
Deterministic key sorting. Forces lexicographical sorting on multi-page queries, preventing duplicate item shifts across pagination intervals.

Concurrency scaling limits

Throughput gains flatten beyond eight concurrent sessions. Two primary physical constraints set this ceiling:

Authentication handshakes: Establishing initial session credentials requires 6 to 9 seconds per thread, expanding to 14 seconds when more than 12 sessions negotiate simultaneously.
Serial work chunk size: The single longest query partition requires 8 seconds to fetch and cannot be subdivided further, forming an absolute execution floor.

Deploying four to six worker sessions captures 90% of maximum throughput while preserving minimal memory overhead.

Reproduction command and environment

python -m benchmark.runner --workload sweeps --pool 4 --output-dir ./results
python -m benchmark.telemetry --input ./results/telemetry.json --export-svg