Benchmark of Java Virtual Threads vs WebFlux

This Java project benchmarks a simple Spring Boot 4.1 microservice using configurable scenarios, comparing Java Virtual Threads (introduced by Project Loom, JEP 444) using Tomcat and Netty with Spring WebFlux (relying on Project Reactor) using Netty.

All benchmark results below come from a dedicated bare metal Ubuntu 24.04 test environment using Java 25 unless specified otherwise. The benchmark also runs monthly on GitHub-hosted runners, using all combinations of (Ubuntu 22.04, Ubuntu 24.04) and (Java 21, Java 25).

Background

Both Spring WebFlux and Virtual Threads are alternative technologies to create Java microservices that support a high number of concurrent users, mapping all incoming requests to very few shared operating system threads. This reduces the resource overhead incurred by dedicating a single operating system thread to each user.

Spring WebFlux was first introduced in September 2017. Virtual Threads were first introduced as preview feature with Java 19 and were fully rolled out with Java 21 in September 2023.

TL;DR

Note

Headline findings from the latest run (Amazon Corretto 25.0.3, bare metal, June 2026; raw data in results/):

Virtual Threads on Netty (using blocking code) matched or beat WebFlux on Netty (using non-blocking code relying on Mono and Flux from Project Reactor) across most - but not all - scenarios and metrics:

Counting every per-metric, per-scenario contest between the two Netty approaches on the default scenarios, Virtual Threads on Netty won ~50%, WebFlux on Netty won ~26%, and ~25% had no clear winner.
Adding Virtual Threads on Tomcat back into the comparison, the overall win-count ranking on the default scenarios is Virtual Threads on Netty (38% of contests), Virtual Threads on Tomcat (22%), WebFlux on Netty (17%), with 22% undecided.
In the heaviest fixed-rate get-movies scenarios (40k and 60k simultaneous connections), Virtual Threads on Netty held a flat ~100 ms median latency, while WebFlux on Netty degraded sharply at 60k (median ~1.2 s, P99 ~3.3 s, ~99% heap, ~108 s of GC time) and served ~31% fewer requests.
The advantage is not universal. In the lighter fixed-rate get-time scenarios the two Netty approaches track each other closely, and at 60k get-time WebFlux on Netty actually delivered slightly lower tail latency (P90/P99 222/613 ms vs 288/734 ms) and marginally more requests.
Both Netty approaches accepted and served every fixed-rate connection burst up to 60k connections with zero errors.

Virtual Threads on Tomcat are not recommended for bursty, high-connection workloads:

Higher steady-state resource use than either Netty approach where all three pass - e.g. at 10k fixed-rate get-movies, ~42% average CPU and ~53% average heap versus ~32–34% CPU and ~37–38% heap for the Netty approaches.
It fails the fixed-rate scenarios once the simultaneous-connection burst exceeds its acceptor's capacity - from 20k connections upward for get-movies - with >99% connection refused errors, even though average CPU during those runs stays well below saturation. The bottleneck is TCP connection acceptance, not request processing. Both Netty approaches accept the identical bursts with zero errors. This even happens in one default scenario.
Tomcat does serve high user counts when the load is paced: it passes every paced 60k spike scenario cleanly, because client think-time keeps only a few thousand connections open at once. See High Load Results.

Benchmark Winners

Below are top-performing approaches across all scenarios and metrics, visualizing the contents of results/scenarios-default/results.csv:

Each cell shows the metric values of best approach (on top) and runner-up.
- What "best" is depends on the metric: A lower value is better for all metrics except for metrics starting with requests_ok, requests_per_second, or sockets.
- Approaches which encountered request errors are ranked below those approaches with only successful requests. Such "failed" approaches have all their metric values printed in red and suffixed with E.
- An overall ranking based on the win count of each approach is shown in the legend: (1) indicates the overall best approach, (2) the runner-up, and so on. This overall ranking is also shown next to each metric value.
Cells are colored based on the winning approach. The darker the color, the bigger the lead of the winning approach over the runner-up. If cells are white or faded, there's no clear winner as the top two approaches performed similarly.
For detailed charts on each approach and test scenario combination, have a look at the second half of this document.
All measurements below were performed on the dedicated, non-virtualized test environment described under Results. Additional monthly measurements are performed on virtualized GitHub-hosted Runners.

All Approaches

This chart compares Project Loom (on both Tomcat and Netty) with Project Reactor (on Netty). Across the 26 metrics × 11 default scenarios (286 per-cell contests), the win-count ranking is Virtual Threads on Netty (38%), Virtual Threads on Tomcat (22%), WebFlux on Netty (17%), with the remaining 22% undecided.

Virtual Threads on Tomcat edges out WebFlux on Netty here on win count despite its catastrophic failure in 10k-vus-and-rps-get-movies-call-depth-1 (where the error-derank rule costs it every metric), largely because it records the lowest GC counts of the three - winning both GC columns (count and time) in 8 of the 11 scenarios - and the lowest P99 latency in all five 20k spike scenarios. Note that this win-count ranking weights all 26 metrics equally; on the latency/throughput metrics that matter most for a service, the two Netty approaches lead.

Netty-based Approaches

This chart is based on the same benchmark as before, but only considers the two Netty-based approaches. Head-to-head, Virtual Threads on Netty wins ~50% of the per-cell contests, WebFlux on Netty ~26%, and ~25% show no clear winner.

Benchmark Features

Fully automated and CLI-driven by running a single command: benchmark.sh.
Different test scenario files, each containing one or more scenarios. Example: src/main/resources/scenarios/scenarios-default.csv.
Operating system thread re-use by waiting and by performing transitive HTTP calls of configurable call depth.
Interacts with realistic JSON APIs.
Creates single PNG image via Matplotlib for each combination of scenario and approach which contains:
- Raw latencies and P50/90/99 percentiles, as well as any errors.
- System metrics for CPU, RAM, sockets, and network throughput.
- JVM metrics such as heap usage, garbage collections (GCs), and platform thread count.
Creates summary PNG image of all scenarios which shows best approaches.

Benchmark Design

The benchmark is driven by k6 which repeatedly issues HTTP requests to a service listening at http://localhost:8080/

The service exposes multiple REST endpoints. The implementation of each has the same 3 stages:

HTTP Call: If $delayCallDepth > 0, call GET /$approach/epoch-millis recursively $delayCallDepth times to mimic calls to upstream service(s).
- By default, all approaches use Spring Boot's WebFlux WebClient based on Netty.
- The scenarios in scenarios-clients.csv compare the WebClient with Spring Boot's RestClient using various client implementations. For details see the Multi-Client Scenarios chapter.
Wait: If $delayCallDepth = 0, wait $delayInMillis (default: 100) to mimic the delay incurred by a network call, filesystem access, or similar.
- Whilst the request waits, its operating system thread can be reused by another request.
- The imperative approaches (platform-tomcat, loom-tomcat, and loom-netty) use blocking wait whilst the reactive approach (webflux-netty) uses non-blocking wait.
Calculate and Return Response specific to REST endpoint.

Sample Flow

Get all movies using loom-netty approach, an HTTP call depth of 1 and a delay of 100 milliseconds:

sequenceDiagram
    participant k6s
    participant service
    k6s->>+service: GET /loom-netty/movies?delayCallDepth=1&delayInMillis=100
    service->>+service: GET /loom-netty/epoch-millis?delayCallDepth=0&delayInMillis=100
    service->>service: Wait 100 milliseconds
    service-->>-service: Return current epoch millis
    service->>service: Find movies
    service-->>-k6s: Return movies

REST APIs

The microservice under test exposes several RESTful APIs. In the following descriptions, $approach is the approach under test and can be one of loom-tomcat, loom-netty, and webflux-netty.

All REST APIs support the following query parameters:

delayCallDepth: Depth of recursive HTTP call stack to $approach/epoch-millis endpoint prior to server-side delay; see Scenario Columns for more details.
delayInMillis: Server-side delay in milliseconds; see Scenario Columns for more details.

epoch-millis

The TimeController returns the milliseconds since the epoch, i.e. 1 Jan 1970:

This is one of the simplest possible APIs to provide a best-case performance scenario.
Supported requests:
- GET /$approach/epoch-millis

movies

The MovieController gets and saves movies which are stored in an H2 in-memory DB via Spring Data JPA, fronted by a Caffeine-backed Spring Boot cache:

This is a realistic JSON API as exposed by a typical microservice.
Several hard-coded movies by three directors are provided.

DB Considerations:

By default, writes are not saved since the code under test is identical for all approaches and would thus only contribute to CPU use. However, this can be controlled with the Spring Boot property loom-webflux.repo-read-only in src/main/resources/application.yaml.
The H2 DB was chosen for the same reason. To swap it for PostgreSQL, specify postgres in the serverProfiles column of the scenario CSV file. See scenarios-postgres.csv and PostgreSQL results.

Supported requests:

GET /$approach/movies?directorLastName={director}:
- Returns movies by the specified director.
- Supported {director} values and their respective response body size in bytes, based on the default movies:
  - Allen: 1664 bytes (unindented)
  - Hitchcock: 1622 bytes (unindented)
  - Kubrick: 1235 bytes (unindented)
POST /$approach/movies:
- Saves one or more movies.
- The request body is the sample movies file (7,010 bytes on disk). Before POSTing, k6 re-serialises it to compact JSON via JSON.stringify(JSON.parse(...)), so the body actually sent over the wire is ~4.1 KB.

Requirements

Software

Unix-based OS; tested with Ubuntu 22.04 and 24.04
Java 21 or above
k6 to drive load and measure latency as well as throughput.
Python 3 with Matplotlib, pandas, and NumPy to drive load and generate benchmark comparison reports.
sar/sadf to measure system resource use
Python 3 and Matplotlib to convert latency and system CSV measurements into a PNG image

Hardware

The hardware requirements depend on the scenarios you intend to run. If you run all scenarios (e.g. via ./benchmarks.sh), then the following is recommended:

CPU: Intel i5 12600K or similar
RAM: 32 GiB

If you only run the default scenarios configured in src/main/resources/scenarios/scenarios-default.csv, the following is sufficient:

CPU: Intel 6700K or similar
RAM: 16 GiB

Setup

The following instructions assume you are using a Debian-based Linux such as Ubuntu 22.04 or 24.04.

Java 21

You'll need Java 21 or above:

sudo apt install openjdk-21-jdk

k6

k6 is used to load the service:

sudo gpg -k
sudo gpg --no-default-keyring --keyring /usr/share/keyrings/k6-archive-keyring.gpg --keyserver hkp://keyserver.ubuntu.com:80 --recv-keys C5AD17C747E3415A3642D57D77C6C491D6AC1D69
echo "deb [signed-by=/usr/share/keyrings/k6-archive-keyring.gpg] https://dl.k6.io/deb stable main" | sudo tee /etc/apt/sources.list.d/k6.list
sudo apt-get update
sudo apt-get install k6

Python 3, matplotlib, sar and sadf

Python 3 and matplotlib are used to convert the CSV output of k6 and sar/sadf to a single PNG chart. The sar and sadf tools come as part of sysstat and are used to measure resource use. To install them run:

sudo apt update && sudo apt install -y python3 python3-matplotlib python3-pandas python3-numpy sysstat

Linux Host Tuning

The benchmark uses high numbers of direct HTTP client connections, so the host must be tuned consistently with the server connection limits used by the application. These settings are intended for benchmark hosts, not as general production-server guidance.

The executable source of truth is src/main/bash/tune-benchmark-host.sh. Local ./benchmark.sh executions check these values before building and request sudo only when a setting needs to be changed:

./src/main/bash/tune-benchmark-host.sh

Use --check to verify the current runtime values without applying changes:

./src/main/bash/tune-benchmark-host.sh --check

The benchmark sysctl baseline is:

sudo tee -a /etc/sysctl.conf <<'EOF'
net.ipv4.ip_local_port_range = 1024 65535
net.ipv4.tcp_timestamps = 1
net.ipv4.tcp_tw_reuse = 2
net.ipv4.tcp_max_syn_backlog = 65535
net.core.somaxconn = 65535
fs.file-max = 1048576
EOF
sudo sysctl -p

ip_local_port_range supports high local k6 concurrency, tcp_timestamps is required for TCP TIME-WAIT reuse semantics, and tcp_tw_reuse = 2 limits reuse to loopback-focused benchmark traffic on modern Linux. tcp_max_syn_backlog and somaxconn avoid measuring connection queue limits during ramp-up. The effective server accept backlog is still the lower of the application's listen backlog and net.core.somaxconn.

CI benchmark runs use non-interactive sudo:

sudo -n ./src/main/bash/tune-benchmark-host.sh

CI hosts must either provide non-interactive sudo for these sysctl writes or preconfigure the same runtime values before running the benchmark. CI must not wait for a sudo password prompt.

File Descriptor Limits

fs.file-max is a system-wide ceiling. It does not replace per-process open-file limits, so both k6 and the JVM process need a sufficient nofile limit:

printf '* soft nofile 1048576\n* hard nofile 1048576\n' | sudo tee -a /etc/security/limits.conf

If the benchmark service is launched through systemd, configure an equivalent LimitNOFILE. To verify active limits:

ulimit -n
cat /proc/$(pgrep -n k6)/limits | grep "Max open files"
cat /proc/$(pgrep -n java)/limits | grep "Max open files"

Log out and back in after changing persistent login limits.

Execution

benchmark.sh

Run a benchmark for each combination of approaches and scenarios defined in a scenario CSV file. Results are stored in build/results/. The script checks benchmark host tuning before the run and may request sudo locally if runtime sysctl values need to be applied:

./benchmark.sh

Usage as per benchmark.sh -h:

Usage: benchmark.sh [OPTION]... [SCENARIO_FILE]
Runs benchmarks configured by a scenario file.

SCENARIO_FILE:     Scenario configuration CSV file in src/main/resources/scenarios/. Default: scenarios-default.csv

OPTION:
  -a <approaches>  Comma-separated list of approaches to test. Default: loom-tomcat, loom-netty, webflux-netty
                   Supported approaches: platform-tomcat, loom-tomcat, loom-netty, webflux-netty
  -C               Keep CSV files used to create chart. Default: false
  -w <seconds>     Override warmupDurationInSeconds from the scenario CSV for all scenarios. Default: value from CSV
  -t <seconds>     Override testDurationInSeconds from the scenario CSV for all scenarios. Default: value from CSV
  -h               Print this help

benchmarks.sh

This is a wrapper over benchmark.sh and supports multiple scenario files:

./benchmarks.sh

Usage as per benchmarks.sh -h:

Usage: benchmarks.sh [OPTION]... [SCENARIO_FILE]...
Wrapper over benchmark.sh that supports multiple scenario files and optionally suspends the system on completion.

SCENARIO_FILE:           Zero or more space-separated scenario configuration CSV files in src/main/resources/scenarios/.
                         Default: scenarios-default.csv scenarios-clients.csv scenarios-deep-call-stack.csv scenarios-postgres.csv scenarios-sharp-spikes.csv scenarios-soaktest.csv scenarios-high-load.csv

OPTION:
  -d, --dry-run          Print what would be done without actually performing it.
  -k, --kill-java        Kill all Java processes after each benchmark. Default: false
  -o, --options "<opts>" Pass additional options to the benchmark.sh script. Run "./benchmark.sh -h" for supported options.
  -s, --suspend          Suspend the system upon completion of the script. Default: false
  -h, --help             Show this help message and exit.

Please note that the default configured scenarios may take several hours to complete.

Approaches

platform-tomcat: Platform threads using Tomcat server
loom-tomcat: Virtual Threads using Tomcat server
loom-netty: Virtual Threads on Netty server
webflux-netty: WebFlux on Netty server

All approaches use the same Spring Boot 4.1.0 version.

Scenarios

Default Scenarios

These 11 scenarios cover a mixture of load patterns, from a 5-user smoke test and a 1k no-delay baseline up to 20k users, mixing fixed-rate (*-vus-and-rps-*) and paced (*-spike-*) shapes, plain get/get-post workloads, and recursive call depths of 0–2.

Config: ./src/main/resources/scenarios/scenarios-default.csv
Results: ./results/scenarios-default/results.md

High-Load Scenarios

These use two deliberately different load shapes - do not read them as one continuous user-count scale:

Fixed-rate scenarios (*-vus-and-rps-*) set connections = requestsPerSecond with no client think-time, so all N virtual users open and continuously drive their connections at once (~N simultaneous connections from the first second). They are swept across 10k → 20k → 40k → 60k users to locate each endpoint's connection-acceptance threshold.
Paced scenarios (*-spike-*) ramp the virtual-user count up and down with 1–3s of think-time between requests and no fixed request rate, so connections are established gradually and, because each user spends most of its time sleeping, only a few thousand are ever in flight at once. They run at 60k users.

As a result, a paced 60k scenario imposes far less instantaneous connection-establishment pressure than a fixed-rate scenario at the same - or even a much lower - user count. See High Load Results for why this matters.

Config: ./src/main/resources/scenarios/scenarios-high-load.csv
Results: ./results/scenarios-high-load/results.md

Multi-Client Scenarios

These scenarios compare both Spring Boot RestClient and WebClient implementations with each other.

Config: ./src/main/resources/scenarios/scenarios-clients.csv
Results: ./results/scenarios-clients/results.md

All scenarios except for those tested with a webflux-netty approach use the WebClient or RestClient implementation specified in the scenario name. However, the webflux-netty approach always uses a fully reactive approach and therefore always uses the non-blocking WebClient.

The following clients are compared:

Spring Boot RestClient based on:
Spring Boot WebClient based on:
- Netty

Other Scenarios

deep-call-stack: High delay call depths
postgres: Use PostgreSQL (started via Docker) instead of H2
sharp-spikes: Intermittent sharp load spikes from 0 to 10/20/30k users
soaktest: Slow ramp-up to 10k users over ~16 minutes, followed by a ~4-minute ramp-down (20-minute test)

Steps

The benchmark run for each $scenario consists of the following phases and steps:

Before Benchmark

Build and start the Spring Boot service with a specific $approach as Spring Boot profile, using the config in src/main/resources/application.yaml and overridden by src/main/resources/application-$approach.yaml if defined.

Benchmark

Run the benchmark as configured by the $scenario.
For each $resultType (i.e. latency, system, or jvm), create a CSV file at build/results/$scenario/$approach-$resultType.csv.

After Benchmark

Convert CSV files into build/results/$scenario/$approach.png
Delete the CSV files unless the -C CLI option was specified.
Stop the service.

Config

Common

The build.gradle.kts file fixes the heap to 2 GiB (-Xms2g -Xmx2g) for the benchmarked service.
The src/main/resources/application.yaml file enables HTTP/2.
Time-outs are 60s: the WebClient response-timeout (used for the service's upstream calls) and the Hikari DB connectionTimeout.

Scenario-specific

Each line in src/main/resources/scenarios/scenarios-default.csv configures a test scenario which is performed first for Java Virtual Threads, then for WebFlux.

Example

scenario	k6Config	serverProfiles	delayCallDepth	delayInMillis	connections	requestsPerSecond	warmupDurationInSeconds	testDurationInSeconds
5k-vus-and-rps-get-time	get-time.js		0	100	5000	5000	10	180
20k-vus-smooth-spike-get-movies	get-movies-smooth-vus-spike.js		0	100	20000		0	180

Columns

scenario: Name of scenario. Is printed on top of each diagram.
k6Config: Name of the K6 Config File which is assumed to be in the config folder
serverProfiles: Pipe-delimited Spring profiles which are also used to start and stop Docker containers. For example, specifying the value postgres|no-cache has these effects:
- The Spring Boot profiles postgres,no-cache are added to the default Spring Boot profile of $approach.
- The files src/main/docker/docker-compose-postgres.yaml and src/main/docker/docker-compose-no-cache.yaml (if existent) are used to start/stop Docker containers before/after each scenario run.
delayCallDepth: Depth of recursive HTTP call stack to $approach/epoch-millis endpoint prior to server-side delay.
- Mimics calls to upstream services which allow for reuse of the current platform thread.
- For example, a value of 0 means that the service waits for $delayInMillis milliseconds immediately upon receiving a request.
- Otherwise, it calls the $approach/epoch-millis with ${delayCallDepth - 1}.
- This results in a recursive HTTP-request-based descent into the service, creating a call stack of depth $delayCallDepth.
delayInMillis: Server-side delay of each request, in milliseconds. Mimics a delay such as invoking a DB which allow for reuse of the current platform thread.
connections: Number of TCP connections, i.e. virtual users.
requestsPerSecond: Number of requests per second across all connections. Left empty for scenarios where the number of requests per second is organically derived based on the number of connections, the request latency, and any explicit client-side delays.
warmupDurationInSeconds: Duration of the warm-up iteration before the actual test. Warm-up is skipped if 0.
testDurationInSeconds: Duration of the test iteration.

Test Environment

Unless noted otherwise, all tests were conducted on this test environment.
Preparation: The system was rebooted before each test and quieted down as much as possible. The baseline total CPU use before test start was reduced to a negligible level.
Co-location: Test driver (k6) and server under test (Spring Boot microservice) were co-located on the same physical machine. The aim of this benchmark is not to achieve maximum absolute performance, but rather to compare different server-side approaches with each other. Considering that the test driver and the load it produced was identical for the combination of server-side approach and scenario, this co-location should not affect the validity of the test results.

Hardware

CPU: Intel Core i5-14600K with 14 cores (6 P + 8 E), 20 threads, up to 5.3 GHz, and 125 W base / 181 W turbo power
Motherboard: Asus ProArt Z690-Creator WIFI
RAM: 64GiB DDR5 (2 x Corsair Vengeance 32 GiB 5600 MT/s CL40)
Disk: local NVMe SSD (not on the benchmark hot path: networking is loopback and the database is in-memory H2)
Network: Loopback interface
Virtualization: None; bare metal desktop

Software

OS: Ubuntu 24.04.4 LTS
Kernel: 6.17.0-35-generic
Java: Amazon Corretto 25.0.3.9.1 (build 25.0.3+9-LTS)
Spring Boot 4.1.0

Note

The actual software versions used by a benchmark are automatically determined and shown at the beginning of each results.md file. If there are differences to the above, then the values in results.md are correct.

Results

This chapter shows the results of each default scenario, sorted by ascending scenario load. The results below can also be found in ./results/scenarios-default/results.md.

Errors

During a run, any lines in the client-side or server-side log files which contain the term error (case-insensitive) are preserved in error log files written to the results folder alongside the generated PNG files. (This committed repository keeps only the PNG charts and the aggregated results.csv/results.md; the verbose per-scenario error logs are not committed.) In the latest run, every error originated from loom-tomcat: the committed high-load results.csv records ~37.3 million failed loom-tomcat requests (concentrated in the fixed-rate get-time/get-movies bursts at 20k–60k), almost all of them connection refused as recorded in the verbose, non-committed error logs. The two Netty approaches produced zero errors on this environment.

Any failed requests appear both in the latency chart as red dots, as well as in the RPS chart as part of a continuous orange line. Additionally, they leave a trace in the $approach-latency.csv file, if preserved by running the benchmark with the -C option:

A very small latency below 3ms indicates that the client failed to establish a TCP connection. Example from the latency CSV file: 1715728866471,0.000000,0,dial: i/o timeout,1211. Such requests are not considered when reporting minimum latency since this could obscure the minimum latency of successful requests.
A very large latency at ~60s reflects a request time-out (the server did not respond within the 60s time-out). Example from the latency CSV file: 1715728861008,60001.327066,0,request timeout,1050.

1k-vus-and-rps-get-time-no-delay

A low-load baseline with no server-side delay: 1k virtual users issue 1k requests per second to the epoch-millis endpoint, which returns immediately. With no artificial wait, each request completes in well under a millisecond (median 0.26–0.54 ms across approaches), so this scenario isolates raw request-handling overhead. All three approaches handle it without errors at single-digit average CPU.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

5k-vus-and-rps-get-time

This scenario aims to maintain a steady number of 5k virtual users (VUs, i.e. TCP connections) as well as 5k requests per second (RPS) across all users for 3 minutes:

Each user issues a request and then waits. This wait between consecutive requests is controlled by k6 in order to achieve the desired number of RPS.
The server-side delay is 100ms.
The server returns the current millis since the epoch.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

5k-vus-and-rps-get-movies

Like the previous scenario, but the response body contains a JSON of movies.

For further details, please see the movies section.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

10k-vus-and-rps-get-movies

Like the previous scenario, but 10k virtual users and requests per second.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

10k-vus-and-rps-get-movies-call-depth-1

Like the previous scenario, but mimics a request to an upstream service.

On receiving an incoming HTTP request, the service calls itself via HTTP.
This secondary request then waits 100 milliseconds.

Warning

This is the one default scenario in which loom-tomcat fails: it completed only 16,042 of ~1.74 million requests (a 99.1% error rate, almost entirely connection refused), even though its average CPU stayed at ~20%. The nested self-call doubles the connection demand on Tomcat's connection-per-request acceptor (HTTP keep-alive is deliberately disabled - see Where Tomcat Breaks), and the acceptor cannot keep up. Both Netty approaches serve the full load with zero errors and ~100 ms median latency. This is the same connection-acceptance limit explored in detail under High Load Results.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-stepped-spike-get-movies

This scenario ramps up virtual users (and thus TCP connections) from 0 to 20k in multiple steps, then back down:

Each step has a short riser time when users are increased, followed by a longer tread time when users are held constant.
Each user issues a request, waits for the response, and then waits for a random delay between 1s and 3s. This reduces the load and better mimics real user interactions with a service, assuming the service calls are driven by user interactions with a website that relies on the service under test.
The server-side delay before returning a response is 100ms.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-smooth-spike-get-movies

Like the previous scenario, but linear ramp-up and down.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-smooth-spike-get-post-movies

Like the previous scenario, but instead of just getting movies, we are now additionally saving them:

75% of requests are GET requests which are split into three groups, each requesting movies by a different director.
25% of requests are POST requests.

For further details, please see the movies section.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-smooth-spike-get-post-movies-call-depth-1

Like the previous scenario, but mimics call to upstream service as explained in 10k-vus-and-rps-get-movies-call-depth-1.

Note

For loom-netty and webflux-netty, this scenario became CPU-contended at peak load (CPU max ~90% for both approaches). This raised tail latencies (P99 ~118-119 ms, max ~371-398 ms) but caused no errors.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-smooth-spike-get-post-movies-call-depth-2

Like the previous scenario, but with a recursive call depth of 2 (each request triggers two nested epoch-millis self-calls before the delay). The deeper call stack increases per-request work: versus call depth 1 it raises GC counts by ~1.5× on the Netty approaches and ~2.3× on Tomcat, and pushes average CPU to ~40–48% across approaches. Because this is a paced spike (1–3s client think-time), all three approaches still complete it without errors; the two Netty approaches keep P99 latency at ~119–122 ms.

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

High Load Results

This chapter highlights interesting results based on scenarios-high-load.csv which scales up to 60k users. For full results, see results/scenarios-high-load/results.md.

Summary

Netty Head-to-Head at High Load

Both Netty approaches accept and serve every fixed-rate burst up to 60k connections with zero errors, so the interesting question at high load is which one degrades more gracefully. The answer depends on the workload:

get-movies (a ~1.6 KB JSON response - Allen's movies): Virtual Threads on Netty win clearly at the top end. At 60k fixed-rate get-movies, loom-netty holds a 100 ms median and serves ~11.0M requests, while webflux-netty degrades sharply - its median climbs to ~1.2 s, its P99 to ~3.3 s, heap peaks at ~99%, GC time reaches ~108 s, and it serves ~7.6M requests (~31% fewer). At 40k the two are closer, but loom-netty still has the lower tail latency.
get-time (a tiny response): the two are close, and WebFlux edges ahead at 60k. At 60k fixed-rate get-time, webflux-netty serves marginally more requests (11.07M vs 10.99M) at lower tail latency (P90/P99 222/613 ms vs 288/734 ms) and lower average CPU (61.6% vs 69.3%).

60k fixed-rate scenario	Approach	Requests (ok)	Latency P50/P90/P99/max (ms)	CPU avg %	GC time (ms)
`get-time`	`loom-netty`	10,989,762	100 / 288 / 734 / 3019	69.3	9,266
`get-time`	`webflux-netty`	11,069,032	100 / 222 / 613 / 1954	61.6	8,140
`get-movies`	`loom-netty`	10,971,441	100 / 379 / 757 / 2899	75.5	12,086
`get-movies`	`webflux-netty`	7,572,752	1174 / 2091 / 3319 / 16936	88.4	108,507

In the paced 60k spike scenarios (stepped, smooth, and smooth get-post), the two Netty approaches finish within a fraction of a percent of each other on request count, and their P99 latencies are near-identical on the get-only spikes (within ~1–2 ms). On the mixed get-post workload loom-netty keeps a tighter tail (P99 117 ms vs 137 ms, max 511 ms vs 1107 ms). The takeaway: Virtual Threads on Netty's advantage is real but concentrated in the heaviest fixed-rate response-bound workloads, not a blanket win across every high-load scenario.

Reading the 40k and 60k Results

A result that surprises many readers is that loom-tomcat records almost no successful requests for the higher fixed-rate scenarios with a server-side delay (e.g. 40k), yet passes the paced 60k scenarios cleanly - even though those use more users. This is a consequence of the two load shapes, not of "60k being easier than 40k":

A fixed-rate scenario presents its full user count as ~simultaneous connections (closed-loop, no think-time). The paced scenarios ramp up gradually and keep only a few thousand connections in flight at any instant, because each virtual user sleeps 1–3s between requests. The sockets rows in results.csv confirm this: in the paced 60k scenarios Tomcat holds only ~1.5k sockets on average (peak ~3k), whereas the failing fixed-rate runs only ever momentarily reach the full burst before the acceptor rejects most of it.
loom-tomcat also passes the no-delay fixed-rate scenario at 60k users, with zero errors - and it does so while holding tens of thousands of concurrent sockets (avg ~40k, peak ~60k). So the distinguishing factor is not how many connections stay open. With no server-side delay, requests complete and connections recycle fast enough that Tomcat's acceptor drains the burst and admits the whole 60k - the run still saturates CPU (~94% average) and median latency climbs to ~372 ms under the load, but nothing is refused. Add a 100ms delay and each accepted connection is held for at least that long; the connection-completion rate collapses, the acceptor falls behind the incoming burst, and ~99.7% of it is refused - which is why the failing fixed-rate runs show low average socket counts (most connections never get established).
The fixed-rate failures are TCP connection-acceptance failures (connection refused dominates the error logs), not request-processing failures. The server does not return error responses; its connection acceptor simply cannot drain the accept queue as fast as the connection burst fills it. Latency and CPU during these runs reflect the few requests that did get accepted, not a saturated server.
The two Netty-based approaches accept the identical fixed-rate bursts (including at 40k and 60k) with zero errors. The difference therefore lies in how each server's connection acceptor copes with a large simultaneous-connection burst (Tomcat's thread-based acceptor versus Netty's event-loop acceptor), not in the host or the load generator.

Note

In short, the fixed-rate loom-tomcat failures are a connection-burst acceptance limit specific to the no-think-time shape - not a general inability to serve high user counts (it serves the paced 60k scenarios without errors). The fixed-rate rows are best read as a connection-acceptance stress test; the paced rows are closer to a typical production traffic shape.

All three approaches run near-vanilla server configuration and share the same OS-level tuning (Linux Host Tuning via tune-benchmark-host.sh), which raises connection limits for every contender equally. This result is not an artefact of under-tuning: in separate experiments (runtime overrides, not part of the committed dataset) the 40k loom-tomcat failure persisted - the large majority of connections still refused - even with the host fully tuned (net.core.somaxconn and net.ipv4.tcp_max_syn_backlog at 65535), Tomcat's accept-count raised to 65,000, and HTTP keep-alive enabled, while Netty served the same burst without errors on the same host. The limiting factor is the rate at which Tomcat's thread-based acceptor admits new connections, not the OS accept-queue depth or any single Tomcat setting.

Where Tomcat Breaks, and How to Push It Further

To make the breaking point explicit and reproducible, scenarios-high-load.csv includes a round-number get-movies ladder of fixed-rate scenarios (each with connections = requestsPerSecond, no client think-time) that brackets it for loom-tomcat with this project's configuration:

Scenario	Simultaneous users (= requests/s)	`loom-tomcat` result (latest run)
`10k-vus-and-rps-get-movies`	10,000	passes (0 errors)
`20k-vus-and-rps-get-movies`	20,000	fails (>99.9% `connection refused`; 1,357 of ~3.57M ok)
`40k-vus-and-rps-get-movies`	40,000	fails (~99.7% `connection refused`)
`60k-vus-and-rps-get-movies`	60,000	fails (~99.8% `connection refused`)

For get-movies, the breaking point sits between 10k and 20k simultaneous users, and the cliff is sharp: 10k is clean while 20k is already an almost-total failure (>99.9% refused). The lighter get-time endpoint breaks higher - in this run it passes both 10k and 20k with zero errors and only fails from 40k (~99.3% refused). Both endpoints impose the same 100 ms server-side delay, so the difference comes from per-request cost: get-time returns a few bytes with negligible work, whereas the ~1.6 KB get-movies response (Allen's movies) adds JSON serialisation and I/O per request, leaving the acceptor less headroom and lowering the connection threshold. Both Netty-based approaches pass every one of these fixed-rate points (get-time and get-movies, up to 60k) without errors.

The separate 60k-vus-*-spike scenarios use the same 60,000 users but ramped with 1–3s of think-time, so they are a different workload type - not the next rung of the ladder above - and Tomcat passes them, because pacing keeps only a few thousand connections concurrent (see Reading the 40k and 60k Results).

How to push the threshold higher. This project intentionally runs Tomcat with HTTP keep-alive effectively disabled (server.tomcat.max-keep-alive-requests: 1, keep-alive-timeout: 1s), so every request opens a fresh TCP connection. Re-enabling keep-alive lets clients reuse connections instead of reconnecting on every request, which relieves the acceptor:

server:
  tomcat:
    max-keep-alive-requests: -1   # unlimited (Tomcat's own default is 100); current project value is 1
    keep-alive-timeout: 60s       # keep idle connections open for reuse; current project value is 1s
    accept-count: 65000           # deeper listen backlog to buffer connection bursts

We verified this experimentally with finer-grained probes (runtime overrides, reverted afterwards - the committed configuration is unchanged, and these probe points are not part of the committed dataset): enabling keep-alive raised the failure threshold by a few thousand simultaneous users - a real but limited gain, still short of the committed 20k scenario, which continues to fail under both settings.

The downsides, and why this is not the default.

It does not remove the ceiling. Even with keep-alive enabled, the committed 20k and 40k scenarios still fail. The bottleneck is the rate at which Tomcat's acceptor admits new connections, not connection reuse, so keep-alive only shifts the threshold - it does not fix the burst case. A deeper accept-count / host somaxconn alone made no measurable difference in our tests.
Higher steady-state resource use. Keep-alive pins one open connection (and, under Virtual Threads, one carrier-bound socket and file descriptor) per idle client for the duration of keep-alive-timeout. With many slow or idle clients and a high max-connections, Tomcat holds far more concurrent connections, threads, and memory than the connection-per-request model - the opposite of the lean profile this benchmark otherwise measures.
Comparability. Changing the connection model alters Tomcat's connection-churn characteristics, breaking direct comparison with the historical results in this repository.

For these reasons the benchmark keeps the current, deliberately lean Tomcat setting (a level playing field with near-vanilla per-contender configuration), documents the breaking point above, and leaves the keep-alive trade-off as an informed choice for a developer tuning a real service.

60k-vus-smooth-spike-get-post-movies

Like 20k-vus-smooth-spike-get-post-movies, but scaling up to 60k users.

Name		Name	Last commit message	Last commit date
.github		.github
docs		docs
gradle/wrapper		gradle/wrapper
results		results
src		src
.editorconfig		.editorconfig
.gitignore		.gitignore
AGENTS.md		AGENTS.md
LICENSE		LICENSE
PLANS.md		PLANS.md
README.md		README.md
WORKLOG.md		WORKLOG.md
benchmark.sh		benchmark.sh
benchmarks.sh		benchmarks.sh
build.gradle.kts		build.gradle.kts
compare-benchmarks.sh		compare-benchmarks.sh
gradle.properties		gradle.properties
gradlew		gradlew
gradlew.bat		gradlew.bat
lombok.config		lombok.config
settings.gradle		settings.gradle

Latest commit

History

Folders and files

Repository files navigation

Benchmark of Java Virtual Threads vs WebFlux

Background

TL;DR

Benchmark Winners

All Approaches

Netty-based Approaches

Benchmark Features

Benchmark Design

Sample Flow

REST APIs

epoch-millis

movies

Requirements

Software

Hardware

Setup

Java 21

k6

Python 3, matplotlib, sar and sadf

Linux Host Tuning

File Descriptor Limits

Execution

benchmark.sh

benchmarks.sh

Approaches

Scenarios

Default Scenarios

High-Load Scenarios

Multi-Client Scenarios

Other Scenarios

Steps

Before Benchmark

Benchmark

After Benchmark

Config

Common

Scenario-specific

Example

Columns

Test Environment

Hardware

Software

Results

Errors

1k-vus-and-rps-get-time-no-delay

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

5k-vus-and-rps-get-time

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

5k-vus-and-rps-get-movies

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

10k-vus-and-rps-get-movies

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

10k-vus-and-rps-get-movies-call-depth-1

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-stepped-spike-get-movies

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-smooth-spike-get-movies

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)

20k-vus-smooth-spike-get-post-movies

Virtual Threads (Tomcat)

Virtual Threads (Netty)

WebFlux (Netty)