Sensu Performance Testing

This repository contains the Sensu performance testing assets used to stress test and measure Sensu's capabilities. Performance testing is done for every Sensu major and minor release to help guard against performance regressions.

The Sensu Testbed

The Sensu Testbed is comprised of five bare metal hosts and two gigabit ethernet network switches. Bare metal is used for increased control and consistency between testing runs (single tenant, no hypervisor, etc.). One host is for running thousands of Sensu Agent sessions (A1), three hosts are for running the Sensu Backend cluster (B1, B2, B3), and the final host runs Postgres for the Sensu Enterprise Event Store (P). One of the network switches is used for SSH access to each host and the Sensu Agent sessions traffic to the Backends. The other network switch is used for Sensu Backend etcd and Postgres traffic. The Postgres host uses three 1 gigabit ethernet cards, round-robin bonded (bond0), to increase its network bandwidth.

Hardware

Agents (agents1)

• AMD Ryzen Threadripper 2990WX Processor, 32 Cores, 3.0 GHz, 83MB Cache

• Gigabyte X399 AORUS PRO, DDR4 2666MHz, Triple M.2

• Corsair Vengeance LPX 32GB DDR4 2666MHz CL16 Quad Channel Kit (4x 8GB)

• Intel 660p Series M.2 PCIe 512GB Solid State Drive

• GeForce GT 710, 1GB DDR3

• Cooler Master Wraith Ripper Ryzen ThreadRipper CPU Cooler

• EVGA SuperNOVA 850W Power Supply

• Cooler Master MasterCase H500P Mesh E-ATX Case

Backends (backend1, backend2, backend3)

• AMD Ryzen Threadripper 2920X Processor, 12 Cores, 3.5GHz, 39MB Cache

• Gigabyte X399 AORUS PRO, DDR4 2666MHz, Triple M.2

• Corsair Vengeance LPX 16GB DDR4 2666MHz CL16 Dual Channel Kit (2x 8GB)

• Two Intel 660p Series M.2 PCIe 512GB Solid State Drives

• Intel Gigabit CT PCIe Network Card

• GeForce GT 710, 1GB DDR3

• Noctua NH-U12S TR4-SP3 CPU Cooler

• Corsair CX Series 650W Power Supply

• Corsair Carbide Series 270R Mid Tower ATX Case

Postgres (postgres)

• AMD Ryzen Threadripper 2920X Processor, 12 Cores, 3.5GHz, 39MB Cache

• Gigabyte X399 AORUS PRO, DDR4 2666MHz, Triple M.2

• Corsair Vengeance LPX 16GB DDR4 2666MHz CL16 Dual Channel Kit (2x 8GB)

• Two Intel 660p Series M.2 PCIe 512GB Solid State Drives

• Samsung 970 PRO NVMe M.2 PCIe 1TB Solid State Drive

• Three Intel Gigabit CT PCIe Network Cards

• GeForce GT 710, 1GB DDR3

• Noctua NH-U12S TR4-SP3 CPU Cooler

• Antec Earthwatts EA-500D 500w Power Supply

• Antec Design Sonata Mid Tower ATX Case

Network

• Two Ubiquiti UniFi 8 Port 60W Switches (US-8-60W)

• Eleven Cat 6 5ft Ethernet Cables

General System Tuning

• An Intel 660p 512GB SSD for the root and var partitions

• Disabled TCP syn cookies (net.ipv4.tcp_syncookies = 0)

Postgres Tuning

System

• An Intel 660p 512GB SSD for the Postgres wal, XFS, 4k block size, mounted with noatime and nodiratime

• A Samsung 970 PRO 1TB SSD for the Postgres database, XFS, 4k block size, mounted with noatime and nodiratime

Postgres

max_connections = 200

shared_buffers = 10GB

maintenance_work_mem = 1GB

vacuum_cost_delay = 10ms
vacuum_cost_limit = 10000

bgwriter_delay = 50ms
bgwriter_lru_maxpages = 1000

max_worker_processes = 8
max_parallel_maintenance_workers = 2
max_parallel_workers_per_gather = 2
max_parallel_workers = 8

synchronous_commit = off

wal_sync_method = fdatasync
wal_writer_delay = 5000ms
max_wal_size = 5GB
min_wal_size = 1GB

checkpoint_completion_target = 0.9

autovacuum_naptime = 10s
autovacuum_vacuum_scale_factor = 0.05
autovacuum_analyze_scale_factor = 0.025

Sensu Backend Tuning

System

• An Intel 660p 512GB SSD for the Sensu Backend embedded etcd (wal and data), ext4 (defaults)

Testing Process

The following steps are intended for Sensu Engineering use, they are shared here for transparency.

Wake the Testbed from the SSH jump host:

./wake.sh

Start up Postgres and do some cleanup:

ssh postgres

systemctl start postgresql-11.service

systemctl status postgresql-11.service

psql 'user=sensu password=P@ssw0rd!'

delete from events;

VACUUM FULL;

\q

Wipe Sensu Backends and start them up (do on all three):

ssh backend1

rm -rf /mnt/data/sensu/sensu-backend

systemctl start sensu-backend.service

systemctl status sensu-backend.service

Configure the Enterprise license and Postgres Event Store (from backend1):

ssh backend1

sensuctl configure

sensuctl create -f sensu-perf/license.json

sensuctl create -f sensu-perf/postgres.yml

In either separate SSH sessions, tmux, or screen panes run (from agents1):

ssh agents1

cd sensu-perf/tests/3-backends-20k-agents-4-subs-pg/

./loadit1.sh

./loadit2.sh

./loadit3.sh

./loadit4.sh

The loadit tool must continue to run for the whole duration of the performance test (do not interrupt).

Create Sensu checks that target the newly created Agent sessions (from backend1):

NOTE: It is recommended to create 4 checks at a time, one for each subscription, this gives etcd some time to allocate pages etc. After etcd has had a chance to "warm up", it's generally safe to be more aggresive with check creation.

ssh backend1

cd sensu-perf/tests/3-backends-20k-agents-4-subs-pg/checks

sensuctl create -f check1.yml

sensuctl create -f check2.yml

sensuctl create -f check3.yml

sensuctl create -f check4.yml

Use Grafana to observe system performance. Watch service logs for any red flags (e.g. increased etcd range request times). Do not forget to collect profiles when you observe anomalous behaviour! Use Grafana to compare the test results with previous test runs.

Test Results

The Sensu Testbed

Postgres Event Storage

Using the 3-backends-20k-agents-4-subs-pg assets and configuration, the Sensu Testbed was able to comfortably handle 40,000 Sensu Agent connections (and their keepalives) and process over 36,000 events per second. The testbed could process over 40,000 events per second, however, the cluster would periodically throttle Agent check executions with back pressure.

Embedded Etcd Event Storage

Using the 3-backends-6k-agents-3-subs assets and configuration, the Sensu Testbed was able to comfortably handle 12,000 Sensu Agent connections (and their keepalives) and process over 8,500 events per second.