Agent Mesh
Go a2aa2a-protocolagent-meshagentic-aiinferencemcpmeshmesh-networknetwork-securitynetworking-aip2p-network
scale: size the guest from measurement, not from a guess
Antonio Ojea83e5e0ca63d6
diff --git a/scripts/launch-microvms.sh b/scripts/launch-microvms.shindex eb747e91..a7981906 100755--- a/scripts/launch-microvms.sh+++ b/scripts/launch-microvms.sh@@ -20,7 +20,18 @@ AGENT_DOMAIN="${AGENT_DOMAIN:-scale.sam-mesh.dev}"# Firecracker defaults to 128 MiB, which silently decides how many agents fit# on a host. Saying it out loud makes density a parameter of the experiment# rather than a property of the tool.-VM_MEM_MIB="${VM_MEM_MIB:-256}"+#+# 160 is the smallest size measured to run the example Python harness at full+# speed: 144 works but boots a second slower for being cramped, and 136 never+# reaches the boundary at all. A heavier agent needs more, so measure your own+# rootfs with tests/scale/measure-guest.sh rather than trusting a number+# written for this one.+#+# Worth knowing when budgeting: mem_size_mib is a ceiling, not an allocation.+# The host only pays for pages the guest touches, so raising it is cheap and+# the real cost per agent is roughly the guest's working set plus about 18 MiB+# for its sam-box.+VM_MEM_MIB="${VM_MEM_MIB:-160}"VM_VCPUS="${VM_VCPUS:-1}"if [ ! -d "$WORKDIR" ]; thendiff --git a/tests/e2e/microvm_sandbox.bats b/tests/e2e/microvm_sandbox.batsindex 69913690..28b10541 100644--- a/tests/e2e/microvm_sandbox.bats+++ b/tests/e2e/microvm_sandbox.bats@@ -13,6 +13,10 @@# This deliberately does not stand up a mesh. A mesh would make the test slower# and would not make it prove anything more: what is in question is whether a# microVM can reach the boundary at all.+#+# FC_MEM_MIB sizes the guest, defaulting to 160: the smallest size measured to+# run the example Python harness at full speed. Raise it for a heavier agent,+# and see tests/scale/measure-guest.sh to measure your own.load "lib/container_mesh.bash"@@ -31,7 +35,13 @@ setup() {# Firecracker CI kernels are built without CONFIG_TUN: they carry vsock and# virtio and nothing else. Checked here because the alternative is a guest# that panics with its explanation truncated by the panic itself.- strings "${kernel}" 2>/dev/null | grep -q "Universal TUN/TAP" \+ #+ # grep -c rather than grep -q: where pipefail is set, a -q that matches exits+ # early, strings takes SIGPIPE, and the pipeline reports failure for having+ # succeeded.+ local tun_driver+ tun_driver="$(strings "${kernel}" 2>/dev/null | grep -c "Universal TUN/TAP" || true)"+ [[ "${tun_driver}" -gt 0 ]] \|| skip "guest kernel has no TUN driver; agent sandboxes need CONFIG_TUN"FC_DIR="$(mktemp -d /tmp/fc-smoke-XXXXXX)"@@ -79,7 +89,7 @@ fc_put() {\"kernel_image_path\": \"${kernel}\",\"boot_args\": \"console=ttyS0 reboot=k panic=1 pci=off\"}"- fc_put /machine-config '{ "vcpu_count": 1, "mem_size_mib": 256 }'+ fc_put /machine-config "{ \"vcpu_count\": ${FC_VCPUS:-1}, \"mem_size_mib\": ${FC_MEM_MIB:-160} }"fc_put /drives/rootfs "{\"drive_id\": \"rootfs\",\"path_on_host\": \"${FC_DIR}/rootfs.ext4\",diff --git a/tests/scale/measure-guest.sh b/tests/scale/measure-guest.shnew file mode 100755index 00000000..c754e056--- /dev/null+++ b/tests/scale/measure-guest.sh@@ -0,0 +1,112 @@+#!/usr/bin/env bash+#+# Copyright 2026 Google LLC+#+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+# http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.+#+# What does one agent sandbox actually need?+#+# How many agents fit on a host is decided by a number somebody has to choose,+# and choosing it by guesswork is how a scale run ends up either wasting most of+# a machine or dying halfway through. This measures it: the guest is booted at+# decreasing memory sizes and asked whether an agent still gets as far as+# opening a flow to the boundary.+#+# It reports three things per size. Whether it worked at all, how long the guest+# took from start to that first flow -- a size that technically works but boots+# slowly is a size under pressure -- and what the microVM cost the host, which+# is the figure density is actually built from.+#+# The numbers describe whatever agent is in the rootfs. Re-run it against your+# own rather than trusting a table written for someone else's.+#+# Usage:+# FC_KERNEL=... FC_ROOTFS=... tests/scale/measure-guest.sh+# SIZES="256 192 160" tests/scale/measure-guest.sh++set -uo pipefail++KERNEL="${FC_KERNEL:-/opt/microvm/vmlinux.bin}"+ROOTFS="${FC_ROOTFS:-/opt/microvm/rootfs.ext4}"+FC="${FC_BIN:-firecracker}"+SIZES="${SIZES:-512 384 256 192 160 144 136 128}"++command -v "${FC}" >/dev/null 2>&1 || { echo "firecracker not found; set FC_BIN" >&2; exit 1; }+[[ -f "${KERNEL}" ]] || { echo "no kernel at ${KERNEL}; set FC_KERNEL" >&2; exit 1; }+[[ -f "${ROOTFS}" ]] || { echo "no rootfs at ${ROOTFS}; set FC_ROOTFS" >&2; exit 1; }++# An agent sandbox has no network device and reaches the mesh through a tun, so+# a kernel without the driver gives it no route and this measures nothing.+#+# grep -c rather than grep -q: under pipefail, a -q that matches exits early,+# strings takes SIGPIPE, and the pipeline reports failure for having succeeded.+tun_driver="$(strings "${KERNEL}" 2>/dev/null | grep -c "Universal TUN/TAP" || true)"+if [[ "${tun_driver}" -eq 0 ]]; then+ echo "guest kernel has no TUN driver; agent sandboxes need CONFIG_TUN" >&2+ exit 1+fi++printf '%-9s %-9s %-14s %s\n' "mem_mib" "reached" "boot_to_flow" "host_rss_mib"++for mem in ${SIZES}; do+ dir="$(mktemp -d /tmp/sam-guest-size-XXXXXX)"+ cp "${ROOTFS}" "${dir}/rootfs.ext4"++ # Stands in for sam-box: the flow only has to arrive, because what is being+ # measured is whether the guest had enough memory to get that far.+ socat -u "UNIX-LISTEN:${dir}/vm.vsock_1080,fork" "CREATE:${dir}/arrived" &+ socat_pid=$!++ "${FC}" --api-sock "${dir}/fc.sock" >"${dir}/fc.log" 2>&1 &+ fc_pid=$!+ for _ in $(seq 1 50); do [[ -S "${dir}/fc.sock" ]] && break; sleep 0.1; done++ put() {+ curl -sf -X PUT --unix-socket "${dir}/fc.sock" "http://localhost$1" \+ -H 'Content-Type: application/json' -d "$2" >/dev/null+ }++ put /boot-source "{\"kernel_image_path\":\"${KERNEL}\",\"boot_args\":\"console=ttyS0 reboot=k panic=1 pci=off\"}"+ put /machine-config "{\"vcpu_count\":${VM_VCPUS:-1},\"mem_size_mib\":${mem}}"+ put /drives/rootfs "{\"drive_id\":\"rootfs\",\"path_on_host\":\"${dir}/rootfs.ext4\",\"is_root_device\":true,\"is_read_only\":false}"+ put /vsock "{\"vsock_id\":\"vsock0\",\"guest_cid\":3,\"uds_path\":\"${dir}/vm.vsock\"}"++ started="$(date +%s%N)"+ put /actions '{"action_type":"InstanceStart"}'++ reached="no"+ elapsed="-"+ for _ in $(seq 1 120); do+ if [[ -s "${dir}/arrived" ]]; then+ reached="yes"+ elapsed="$((($(date +%s%N) - started) / 1000000))ms"+ break+ fi+ # A guest that panicked is not going to recover; stop waiting on it.+ kill -0 "${fc_pid}" 2>/dev/null || break+ sleep 0.25+ done++ # Resident, not configured: the host pays for pages the guest touched, so+ # this is what density is built from rather than mem_size_mib.+ rss="-"+ if kill -0 "${fc_pid}" 2>/dev/null; then+ rss="$(awk '/VmRSS/ {printf "%.0f", $2/1024}' "/proc/${fc_pid}/status" 2>/dev/null || echo -)"+ fi++ printf '%-9s %-9s %-14s %s\n' "${mem}" "${reached}" "${elapsed}" "${rss}"++ kill "${fc_pid}" "${socat_pid}" 2>/dev/null+ wait "${fc_pid}" 2>/dev/null+ rm -rf "${dir}"+done