fix: reserve ephemeral-port headroom for the auto max-connections default

[nvzhihanj]

Problem

client.max_connections defaults to -1, which resolves to the entire
ephemeral-port range (system_maximum_ports × distinct_endpoints). When
connections are established in a burst — e.g. max_throughput at start-up — the
pool tries to bind every ephemeral port at once. The OS holds recently-closed
ports in TIME_WAIT, so the tail of bind() calls intermittently fails with
EADDRNOTAVAIL ("Cannot assign requested address"). The run recovers, but only
after a burst of dropped/retried connections; it's worst when the client shares a
host with other socket activity.

A connection pool should never try to claim 100% of the ephemeral range — that
leaves nothing for TIME_WAIT recycling, the OS, or other processes on the box.

Fix

Resolve the auto (-1) limit to a fraction of the budget instead of the whole
range, so the zero-config default reserves headroom:

• AUTO_MAX_CONNECTIONS_BUDGET_FRACTION = 0.5 — auto claims half the per-endpoint
  budget. That is still a large pool (thousands of connections per endpoint) and
  continues to scale with the number of distinct endpoints.
• An explicit max_connections is unchanged and may still use the full
  budget (the over-budget guard is unchanged).

Also removes the now-dead get_ephemeral_port_limit() / get_used_port_count()
helpers — the static reserve replaces them without reading the racy,
process-global live socket count.

Why this default

The client should work well without the user having to reason about ephemeral
ports or hand-tune max_connections. A static reserve keeps that property: the
default is safe on a busy/co-located host, while workloads that genuinely need
more concurrency add endpoints (the budget scales per distinct destination) or
set an explicit value. Empirically, claiming the full range reliably produced
EADDRNOTAVAIL under a bursty max_throughput start-up, whereas reserving
headroom eliminated it with no change to server-side batch or throughput.

Tests

Updated TestEndpointBudgetScaling for the reserved-headroom default and added a
case asserting the auto limit leaves headroom while an explicit value may use the
full budget.

---

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[gemini-code-assist]

Code Review

This pull request updates the automatic connection limit calculation to only claim a fraction (50%) of the ephemeral-port budget, leaving headroom for the OS to prevent connection establishment failures. Additionally, unused port-counting utility functions have been removed, and the unit tests have been updated and expanded to verify the new behavior. There are no review comments, and I have no additional feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.