As we and our collaborators prepare Round 9 of our Framework Benchmarks project, we had an epiphany:

With high-performance software, a single modern server processes over 1 million HTTP requests per second.

Five months ago, Google talked about load-balancing to achieve 1 million requests per second. We understand their excitement is about the performance of their load balancer1. Part of what we do is performance consulting—so we are routinely deep in request-per-second data—and we recognized a million requests per second as an impressive milestone.

But fast-forward to today, where we see the same response rate from a single server. We had been working with virtual servers and our modest workstations for so long that these data were a bit of a surprise.

The mind immediately begins painting a world of utter simplicity, where our applications' scores of virtual servers are rendered obsolete. Especially poignant is the reduced architectural complexity that an application can reap if its performance requirement can be satisfied by a single server. You probably still want at least two servers for resilience, but even after accounting for resilience, your architectural complexity will likely remain simpler than with hundreds of instances.

Our project's new hardware

For Round 9 of our benchmarks project, Peak Hosting has generously provided us with a number of Dell R720xd servers each powered by dual Xeon E5-2660 v2 CPUs and 10-gigabit Ethernet. Loaded up with disks, these servers are around $8,000 a piece direct from Dell. Not cheap.

But check out what they can do:

techempower@lg01:~$ wrk -d 30 -c 256 -t 40 Running 30s test @ 40 threads and 256 connections Thread Stats Avg Stdev Max +/- Stdev Latency 247.05us 3.52ms 624.37ms 99.90% Req/Sec 27.89k 6.24k 50.22k 71.15% 31173283 requests in 29.99s, 3.83GB read Socket errors: connect 0, read 0, write 0, timeout 9 Requests/sec: 1039305.27 Transfer/sec: 130.83MB

This is output from Wrk testing a single server running Undertow using conditions similar to Google's test (1-byte response body, no HTTP pipelining, no special request headers). 1.039 million requests per second.

Obviously there are myriad variables that make direct comparison to Google's achievement an impossibility. Nevertheless, achieving a million HTTP requests per second over a network without pipelining to a single server says something about the capacity of modern hardware.

It's possible even higher numbers would be reported had we tested a purpose-built static web server such as nginx. Undertow is the lightweight Java web application server used in WildFly. It just happens to be quite quick at HTTP. Here's the code we used for this test:

public class ByteHandler implements HttpHandler { private static final String aByte = "a"; @Override public void handleRequest(HttpServerExchange exchange) throws Exception { exchange.getResponseHeaders().put( Headers.CONTENT_TYPE, TEXT_PLAIN); exchange.getResponseSender().send(aByte); } }

In Round 9 (coming soon, we swear!), you'll be able to see the other test types on Peak's hardware alongside our i7 workstations and the EC2 instances we've tested in all previous rounds. Spoiler: I feel bad for our workstations.

Incidentally, if you think $8,000 is not cheap, you might want to run the monthly numbers on 200 virtual server instances. Yes, on-demand capacity and all the usual upsides of cloud deployments are real. But the simplified system architecture and cost advantage of high-performance options deserve some time in the sun.

1. Not only that, Google expressly said they were not using this exercise to demonstrate the capacity of their instances but rather to showcase their load balancer's performance. However, the scenario they created achieved massive request-per-second scale by load balancing hundreds of instances. We are simply providing a counter-point that the massive scale achieved by hundreds of instances can be trivially mimicked by a single modern server with modern tools. The capacity of a single server may not be surprising to some, but it may come as a surprise to others.

December 17, 2013

Framework Benchmarks Round 8

Merry Christmas web framework performance aficionados! What better way to celebrate the holidays than by cheering on your favorites as they race through a variety of application fundamentals in the biggest web platform grudge match of the season? We certainly can't think of anything more festive.

Now at 90 frameworks and 230 permutations (variations on configuration), Round 8 has something for everyone. And if it doesn't have what you want, you can join the party! We have fruitcake and egg nog. Or maybe not. But we enjoy pull requests; they're almost as good as egg nog.

A veritable rainbow of holiday cheer awaits!

View Round 8 results

View Round 8 results now.

Round 8 notes and observations

  • Go, always the scrappy competitor, flexes some performance muscle and lands a razor-thin victory in JSON serialization on i7 hardware. But be aware, the highest-performance frameworks are network limited in the JSON serialization and Plaintext tests. Anything in this "200k club" is sure to keep overhead at a bare minimum, leaving maximum headroom for your custom application logic.
  • Round 7 was missing some of the Go frameworks due to configuration problems. Those problems have been resolved, and the Go frameworks have returned in Round 8 to reaffirm that Go is a viable performance rival to the JVM.
  • The HipHop PHP VM with no framework and thanks in part to the MySQL driver for PHP, yields dominion over the Updates test. HHVM is impressive in the multiple query test as well. However, hhvm trails plain PHP in the Fortunes test presently. Implementation details may be at play here. If you're interested in testing HHVM with popular PHP frameworks, we would be happy to receive a pull request.
  • Vert.x and Netty have wrestled the Plaintext crown from Undertow, but this rivalry isn't yet settled. Rumor has it they have more improvements in store for Round 9. Meanwhile, a newcomer named Plain (which may rival Go as the most in need of a more search-friendly name; though the irony of Go makes it uncontested champion) is right behind the leaders. Most interestingly, Plain demonstrates the highest Windows performance we've seen by a massive margin (reaching 611,095 pipelined plaintext requests per second on i7). Once again, bear in mind that these tests are network-bound by our gigabit Ethernet.
  • On EC2, the Netty and Vert.x upgrades have paid huge dividends with Netty now breaking 200,000 pipelined plaintext responses per second on a humble m1.large instance.
  • Grizzly performance on JSON serialization is off from its Round 7 showing, but unfortunately, we have not yet determined the cause.
  • The Plaintext test requirements were clarified. It is not necessary to copy the bytes of the small response payload per request. Using a pre-rendered byte buffer for the body is acceptable as long as that is conventional for the platform or framework being tested and response headers are composed normally.
  • The maximum query and update performance for Mongo on EC2 is substantially higher than MySQL. When looking at that chart in particular, consider filtering by your preferred data store to maintain a useful perspective. Related: a late change to the Mongo "schema" intended to replace "id" with "_id" caused some challenges. We further postponed Round 8 to re-run Mongo tests with a schema that provides both columns to allow all tests to complete. We want to normalize the implementations for Round 9.
  • We were targeting early December for Round 8 and we're off by about two weeks. There is still room for improvement toward our goal of a monthly cycle. We will target mid-January for Round 9.


A big thank-you to all of the contributors who have added and improved existing test implementations for Round 8.

The contributors for Round 8 are, in no particular order: @lhotari, @methane, @pseudonom, @lucassp, @aualin, @weltermann17, @kpacha, @nareshv, @martin-g, @bclozel, @ijl, @bbrowning, @sbordet, @purplefox, @stuartwdouglas, @normanmaurer, @kardianos, @hamiltont, and @julienschmidt.

If you have questions, comments, criticism, or would like to contribute a new test or an improvement to an existing one, please join our Google Group or visit the project at Github.

About TechEmpower

We provide web and mobile application development services and are passionate about application performance. Read more about what we do.

October 31, 2013

Framework Benchmarks Round 7

Happy Halloween fans of web development frameworks! After a several-month hiatus, Round 7 of our project measuring the performance of web application frameworks and platforms is available!

View Round 7 results

Round 7 includes many new framework test implementations contributed by the community. They are Falcore, Grizzly, HttpListener, PHPixie, Plain, Racket-WS, Start, Stream, and Treefrog. There are now a whopping 84 frameworks and over 200 individual test permutations.

Many preexisting frameworks' tests have been updated to include more test coverage and/or update dependencies and tune their implementation. To date, the project has processed 344 pull requests from the community. Thanks so much for your contributions. We are grateful for your continued interest!

View Round 7 results now.

Round 7 notes and observations

  • The Round 6 champion Undertow (the web server for WildFly) continues to impress with chart-dominating showings such as 180,000 plaintext requests per second on meager m1.large instances.
  • Thanks to community contributions, the C# tests have been dramatically improved, especially when querying the database. We also have some SQL Server tests in our i7 environment.
  • A contributor prepared scripts for running the benchmark suite on Windows Azure. Unfortunately, we were unable to reach the author of these scripts in the past weeks. If any Azure experts are interested in picking up that work where it exists now, please visit the GitHub repository or the Google Group for the project.
  • The high-performance tier has become significantly more crowded even during this project's relatively short history. Most interesting to us is how many frameworks can easily saturate our gigabit Ethernet with the JSON serialization and plaintext tests, even with our tests' intentionally small payloads. We do not have the hardware necessary to run 10 gigabit Ethernet tests, but if you have a 10 GBE lab and are willing to run the suite, we'd love to publish the results.
  • The benchmark toolset continues to mature gradually, but a lot of room for improvement still exists. A great deal of sanity-checking remains a manual process. If you're a Python programmer and interested in this project, let us know. We have several enhancements we'd like to make to the benchmark tool set (Python scripts), time permitting.
  • This round used a community-review model wherein project participants were able to review preliminary results we were capturing in our i7 environment and submit pull requests. The model is not perfect and will need to improve with each round, but it will help reduce the amount of time we (TechEmpower) need to allocate to each round's sanity checks, meaning quicker turn-around of rounds (see how I spun that as a good thing?).
  • Starting now, we aim to be on a monthly cycle of running official rounds. This helps reduce the perceived severity of configuration problems since they can be addressed in the next run, which is only a month away.
  • We've also pushed the display name for tests into the project, allowing contributors to assign test permutations any name they choose. E.g., "play-scala-anorm" and "aspnet-mvc-mono."
  • One particularly interesting anomaly is the dominance of Windows paired with Mongo on EC2 in the Updates test. The performance is only slightly lower than the same pairing on i7, where in most cases our i7s (2600K workstations, to be precise) and EC2 (m1.large) instances differ by a factor of seven or more. It's possible the Windows EC2 instance is running on a newer host than the Linux EC2 instance, but both are classified as m1.large.
  • Speaking of database tests, in previous rounds, we had used an SSD to host the databases. Prior to finishing Round 7, that SSD failed, so Round 7 is run with ramdisk-backed databases (excluding SQL Server). This project is not a database benchmark so we believed it would be fascinating to see the performance of the full stack when the friction of the database writes is reduced to a bare minimum. As confirmed by our previous spot checking in Round 5, database writes are about 20% to 30% faster across the board when using a ramdisk versus the Samsung 840 Pro SSD we had been using. As expected, reads are unaffected since the tests are designed to allow the database engine to fit the entire data set into memory.


As always, we'd like to say thank you to all of the contributors who have added test implementations for new frameworks or improved existing implementations. Round 7 was unusually long, so we also thank everyone for their patience.

The contributors for Round 7 are numerous. In no particular order: @fernandoacorreia, @kppullin, @MalcolmEvershed, @methane, @KevinHoward, @huntc, @lucassp, @dracony, @weltermann17, @kekekeks, @fwbrasil, @treefrogframework, @yogthos, @oberhamsi, @purplefox, @yz0075, @necaris, @pdonald, @Kepinator, @DavidBadura, @zznate, @nightlyone, @jeapostrophe, @astaxie, @troytoman, @grob, @torhve, @trautonen, @stuartwdouglas, and @xaxaxa. Sincere apologies if we forgot anyone!

If you have questions, comments, criticism, or would like to contribute a new test or an improvement to an existing one, please join our Google Group or visit the project at Github.

About TechEmpower

We provide web and mobile application development services and are passionate about application performance. Read more about what we do.