tech oriented notes to self and lessons learned
TechEmpower Web Framework Benchmarks is a collaborative, open web framework benchmarking project that I blogged about last year (see An open web application framework benchmark). Since the last write-up a new benchmark round has been run with lots of new test implementations. Among others, round 10 benchmark run includes support for the Cassandra NoSQL database, as well as a new Java-based test implementation leveraging Servlet 3 asynchronous processing.
I did some non-scientific analysis of TFB round 10 results and here’re a few observations on the Peak environment, “best” concurrency level results. According to the documentation round 10 and 9 Peak hosting hardware specifications are the same, so the results should be directly comparable.
lwan has overtaken previous run’s #1 performer in the JSON serialization test by a large margin, more than doubling round 9 winner’s throughput. cpoll_cppsp, last round’s winner, also improved its throughput by nearly 700k req/s. Six contestants from round 9 have made it into top 10 in round 10, with all of them being able to improve their performance. OpenResty managed to more than double its round 9 throughput.
My new entrant, servlet3-cass, has taken 5th place in this category, which made it the second best performing Java implementation, losing only slightly to Netty. Based on this, I would conclude that Servlet container overhead doesn’t currently seem to impose a significant bottleneck in this test. It’s interesting to note that there’s about 400 req/s difference between servlet and servlet3-cass performance, though the implementations are very similar. Both have been implemented on the Servlet API, run on the Resin servlet container and use Jackson for JSON processing. The only difference between the two seems to be Jackson JSON library version numbers.
The top 10 test implementations were based on the following programming languages: C, C++, Java and Lua. Of the top 10 performers 5 are Java based, and Ur and Go got dropped out of the top 10 from the previous round.
Since round 9 the #1 performer’s throughput has improved only marginally in this test category. C++ based test implementations dominate the top 4 spots with Lua, Ur and Java also in top 10. Only test implementations based on relational databases made it to top 10, leaving out MongoDB and Cassandra.
Some test implementations seem to have multiple entries in the results table (e.g. cpoll_cppsp-postgres, undertow and activeweb) for this test. I guess this is because the test was run multiple times, but I would’ve expected each framework to be listed only once in the best results table. Also, what’s the exact test methodology like in these cases? Under which conditions does a framework get multiple tries? Some test implementations seem to have experienced performance degradation since round 9. One such example is OpenResty. It would be interesting to analyze the cause further. Was the degradation caused by e.g. changes in the test implementation, infrastructure setup or changes in test methodology?
I was a bit disappointed with the performance of the servlet3-cass test implementation in the database tests. Unfortunately, there’s no resource usage or profiling data available for the test run, and as I don’t have access to a real performance test environment myself, so it’s difficult to analyse potential bottlenecks further.
Because servlet3-cass was able to reach 14 times the single database query throughput with the JSON test, I would reason that no inherent app server framework bottleneck gets reached in the database test. By comparison, the Java and servlet API based test implementation that uses MySQL database, was able to achieve nearly twice the throughput. Two notable implementation differences come to mind between the two test implementations: a) different datastore and b) use of different servlet API (synchronous vs. asynchronous).
Since, Cassandra is in general very fast in reading and writing data by key, there shouldn’t be any inherent reason why it should perform worse than the MySQL based implementation. Some aspects I need to analyse in more detail: 1) Cassandra server configuration 2) Resin servlet container asynchronous processing support and/or configuration 3) check that thread pool size for Resin asynchronous processing and Cassandra driver is optimized for the test server hardware concurrency level. During the test implementation I ran into a couple of bugs in servlet container asynchronous processing support (both Resin + Tomcat. Kudos to both teams for fixing them!), which proved that asynchronous processing is a tricky issue, not only for application, but also app server developers. In order to evaluate whether some aspect of asynchronous processing has a negative performance impact it would be interesting to implement the test using traditional synchronous servlet API.
The best two implementations in this test, start and stream, were able to hold on to their round 9 positions, while both improved throughput by more than 30 %. The top 10 test implementations are based on Dart, Java, C++ and Clojure.
MongoDB based implementations have claimed a triple win in this test with the 4 implementations that follow being PostgreSQL based.
Five new test implementations made it to top 10 compared to the previous round. The top 10 implementations were implemented in C++, Ur and Java. Again, both undertow and undertow edge seem to be included three times in round 9 data table, which looks a bit strange for the best results table.
Having done quite a bit of Perl programming years ago, I was quite fascinated to see an implementation based on the old scripting workhorse make it to top 10.
In this test there’s also been about 10 % performance degradation for the top 5 implementations.
On round 9 the top 10 implementations were all based on MySQL whereas on round 10 three PostgreSQL based implementations have entered the top 10.
The top 10 in this category is occupied by test implementations based on C++, C, Java and Scala. While there’s been only a modest performance increase for the #1 position, the top place has been overtaken by ulib. Four new entrants have made their way into top 10. Netty has gained over 40% throughput increase compared to the previous round.
The TechEmpower team has again put a huge effort into the FrameworkBenchmarks project and done a great job at it! Still, as with everything, there’s always room for improvement and, from a casual test implementation contributor perspective, the following issues should be considered:
I’ve updated a TFB projects results visualization tool I created to also render round 10 results. As before, the tool can be found here:
The project team has been hard at work and they’ve published a tentative release schedule for round 11 results, which are aimed at being released by the end of June 2015. Since round 10, I’ve contributed support for ElasticSearch search server as well as a Scala / Spray based test implementation. Both have been merged into the project repository, so the results should be available when round 11 data gets published. The code for the new test implementation can be found here:
Again, great job TFB development team and contributors – keep up the good work!
Looking forward to seeing round 11 results!
Selecting a platform for your next application development project can be a complex and burdensome undertaking. It can also be very intriguing and a lot of fun. There’s a wide range of different approaches to take: at one end The Architect will attend conferences, purchase and study analyst reports from established technology research companies such as Gartner, and base his evaluation on analyst views. Another approach is to set up a cross-disciplinary evaluation committee that will collect a wishlist of platform requirements from around the organization and make its decision based on a consensus vote. The first approach is very autocratic, while the second can sometimes lead to lack of focus. A clear, coherent vision of requirements and prioritization is essential for the success of the evaluation. Due to these problems, a middle road and a more pragmatic approach is becoming increasingly popular: a tightly-knit group of senior propellerheads use a more empiric method of analysing requirements, study and experiment with potential solution stack elements, brainstorm to produce a short list of candidates to be validated using a hands-on architecture exercises and smell-tests. Though hands-on experimentation can lead to better results, the cost of this method can be prohibitive, so often only a handful of solutions that pass the first phase screening can be evaluated this way.
Platform evaluation criteria depend on the project requirements and may include:
Performance and scalability are often high priority concerns. They are also among those platform properties that can be formulated into quantifiable criteria, though the key challenge here is how to model the user and implement performance tests that accurately model your expected workloads. Benchmarking several different platforms can only add to the cost of benchmarking.
A company called TechEmpower has started a project called TechEmpower Framework Benchmarks, or TFB for short, that aims to compare the performance of different web frameworks. The project publishes benchmark results that application developers can use to make more informed decisions when selecting frameworks. What’s particularly interesting about FrameworkBenchmarks, is that it’s a collaborative effort conducted in an open manner. Development related discussions take place in an online forum and the source code repository is publicly available on GitHub. Doing test implementation development in the open is important for enabling peer review and it allows implementations to evolve and improve over time. The project implements performance tests for a wide variety of frameworks, and chances are that the ones that you’re planning to use are included. If not, you can create your own tests and submit them to be included in the project code base. You can also take the tests and run the benchmarks on your own hardware.
Openly published test implementations are not only useful for producing benchmark data, but can also be used by framework developers to communicate framework performance related best practices to application developers. They also allow framework developers to receive reproducible performance benchmarking feedback and data for optimization purposes.
It’s interesting to note that the test implementations have been designed and built by different groups and individuals, and some may have been more rigorously optimized than others. The benchmarks measure the performance of the framework as much as they measure the test implementation, and in some cases suboptimal test implementation will result in poor overall performance. Framework torchbearers are expected to take their best shot in optimizing the test implementation, so the implementations should eventually converge to optimal solutions given enough active framework pundits.
In the project’s parlance, the combination of programming language, framework and database used is termed “framework permutation” or just permutation, and some test types have been implemented in 100+ different permutations. The different test types include:
Currently, the latest benchmark is Round 9 and the result data is published on the project web page. The data is not available in machine-readable form and it can’t be sorted by column for analysing patterns. It can, however, be imported into a spreadsheet program fairly easily, so I took the data and analyzed it a bit. Some interesting observations could be made just by looking at the raw data. In addition to comparing throughput, it’s also interesting to compare how well frameworks scale. One way of quantifying scalability is to take test implementation throughput figures for the lowest and highest concurrency level (for test types 1, 2, 4 and 6) per framework and plot them on a 2-D plane. A line can then be drawn between these two points with the slope characterizing scalability. Well-scaling test implementations would be expected to have a positive, steep slope for test types 1, 2, 4 and 6 whereas for test types 3 and 5 the slope is expected to be negative.
This model is not entirely without problems since the scalability rating is not relative to the throughput, so e.g. a poorly performing framework can end up having a great scalability rating. As a result, you’d have to look at these figures together.
To better visualize throughput against concurrency level (“Peak Hosting” environment data), I created a small web app that’s available at http://tfb-kippo.rhcloud.com/ (the app is subject to removal without notice).
The JSON serialization test aims to measure framework overhead. One could argue that it’s a bit of a micro benchmark, but it should demonstrate how well the framework does with basic tasks like request routing, JSON serialization and response generation.
The top 10 frameworks were based on the following programming languages: C++, Java, Lua, Ur and Go. C++ based CPPSP was the clear winner while the next 6 contestants were Java -based. No database is used in this test type.
The top 7 frameworks with highest throughput also have the highest scalability rating. After that, both these figures start declining fairly rapidly. This is a very simple test and it’s a bit of a surprise to see such large variation in results. In their commentary TechEmpower attributes some of the differences to how well frameworks work on a NUMA-based system architecture.
Quite many frameworks are Java or JVM based and rather large variations exist even within this group, so clearly neither the language nor the JVM is an impeding factor in this group.
I was surprised about Node.js and HHVM rankings. Unfortunately, the Scala-based Spray test implementation, as well as the JVM-based polyglot framework Vert.x implementation, were removed due to being outdated. Hope to see these included in a future benchmark round.
This test type measures database access throughput and parallelizability. Again, surprisingly large spread in performance can be observed for a fairly trivial test case. This would seem to suggest that framework or database access method overhead contributes significantly to the results. Is the database access technology (DB driver or ORM) a bottleneck? Or is the backend system one? It would be interesting to look at the system activity reports from test runs to analyze potential bottlenecks in more detail.
Before seeing the results I would’ve expected the DB backend to be the bottleneck, but this doesn’t appear to be clear-cut based on the fact that the top, as well as many of the bottom performing test implementations, are using the same DB. It was interesting to note that the top six test implementations use a relational database with the first NoSQL based implementation taking 7th place. This test runs DB read statements by ID, which NoSQL databases should be very good at.
Top performing 10 frameworks were based on Java, C++, Lua and PHP languages and are using MySQL, PostgreSQL and MongoDB databases. Java based Gemini leads with CPPSP being second. Both use MySQL DB. Spring based test implementation performance was a bit of a disappointment.
Where the previous test exercised a single database query per request this test does a variable number of database queries per request. Again, I would’ve assumed this test would measure the backend database performance more than the framework performance, but it seems that framework and database access method overhead can also contribute significantly.
The top two performers in this test are Dart based implementations that use MongoDB.
Top 10 frameworks in this test are based on Dart, Java, Clojure, PHP and C# languages and they use MongoDB and MySQL databases.
This is the most complex test that aims to exercise the full framework stack from request routing through business logic execution, database access, templating and response generation.
Top 10 frameworks are based on C++, Java, Ur, Scala, PHP languages and with the full spectrum of databases being used (MySQL, PostgreSQL and MongoDB).
In addition to reads this test exercises database updates as well.
HHVM wins this test with 3 Node.js based frameworks coming next. Similar to the Single database query test the top 13 implementations work with relational MySQL DB, before NoSQL implementations. This test exercises simple read and write data access by ID which, again, should be one of NoSQL database strong points.
The aim of this test is to measure how well the framework performs under extreme load conditions and massive client parallelism. Since there’s no backend system dependencies involved, this test measures platform and framework concurrency limits. Throughput plateaus or starts degrading with top-performing frameworks in this test before client concurrency level reaches the maximum value, which seems to suggest that a bottleneck is being hit somewhere in the test setup, presumably hardware, OS and/or framework concurrency.
Many frameworks are at their best with concurrency level of 256, except CPPSP which peaks at 1024. CPPSP is the only one of the top-performing implementations that is able to significantly improve its performance as the concurrency level increases from 256, but even with CPPSP throughput actually starts dropping after concurrency level hits the 4,096 mark. Only 12 test implementations are able to exceed 1 M requests per second. Some well-known platforms e.g. Spring did surprisingly poorly.
There seems to be something seriously wrong with HHVM test run as it generates only tens of responses per second with concurrency levels 256 and 1024.
Top 10 frameworks are based on C++, Java, Scala and Lua languages. No database is used in this test.
In the scientific world research must be repeatable, in order to be credible. Similarly, the benchmark test methodology and relevant circumstances should be documented to make the results repeatable and credible. There’re a few details that could be documented to improve repeatability.
The benchmarking project source code doesn’t seem to be tagged. Tagging would be essential for making benchmarks repeatable.
A short description of the hardware and some other test environment parameters is available on the benchmark project web site. However, the environment setup (hardware + software) is expected to change over time, so this information should be documented per round. Also, Linux distribution minor release or the exact Linux kernel version don’t appear to be identified.
Detailed data about what goes on inside the servers could be published, so that externals could analyze benchmark results in a more meaningful way. System activity reports e.g. system resource usage (CPU, memory, IO) can provide valuable clues to possible scalability issues. Also, application, framework, database and other logs can be useful to test implementers.
Resin was chosen as the Java application server over Apache Tomcat and other servlet containers due to performance reasons. While I’m not contesting this statement, but there wasn’t any mention about software versions, and since performance attributes tend to change over time between releases, this premise is not repeatable.
Neither the exact JVM version nor the JVM arguments are documented for JVM based test implementation execution. Default JVM arguments are used if test implementations don’t override the settings. Since the test implementations have very similar execution profiles by definition, it could be beneficial to explicitly configure and share some JVM flags that are commonly used with server-side applications. Also, due to JVM ergonomics different GC parameters can be automatically selected based on underlying server capacity and JVM version. Documenting these parameters per benchmark round would help with repeatability. Perhaps all the middleware software versions could be logged during test execution and the full test run logs could be made available.
Since I’ve worked recently on implementing RESTful services based on JAX-RS 2 API with asynchronous processing (based on Jersey 2 implementation) and Apache Cassandra NoSQL database, I got curious about how this combination would perform against the competition so, I started coding my own test implementation. I decided to drop JAX-RS in this case, however, to eliminate any non-essential abstraction layers that might have a negative impact on performance.
One of the biggest hurdles in getting started with test development was that, at the time I started my project there wasn’t a way to test run platform installation scripts in smaller pieces, and you had to run the full installation, which took a very long time. Fortunately, since then framework installation procedure has been compartmentalized, so it’s possible to install just the framework that you’re developing tests for. Also, recently the project has added support for fully automated development environment setup with Vagrant, which is a great help. Another excellent addition is Travis CI integration that allows test implementation developers to gain additional assurance that their code is working as expected also outside their sandbox. Unfortunately, Travis builds can take a very long time, so you might need to disable some of the tests that you’re not actively working on. The Travis CI environment is also a bit different from the developer and the actual benchmarking environments, so you could bump into issues with Travis builds that don’t occur in the development environment, and vice versa. Travis build failures can sometimes be very obscure and tricky to troubleshoot.
The actual test implementation code is easy enough to develop and test in isolation, outside of the real benchmark environment, but if you’re adding support for new platform components such as databases or testing platform installation scripts, it’s easiest if you have an environment that’s a close replica of the actual benchmarking environment. In this case adding support for a new database involved creating a new DB schema, test data generation and automating database installation and configuration.
Implementing the actual test permutation turned out to be interesting, but surprisingly laborious, as well. I started seeing strange error responses occasionally when benchmarking my test implementation with ab and wrk, especially with higher loads. TFB executes Java based performance implementations in the Resin web container, and after a while of puzzlement about the errors, I decided to test the code in other web containers, namely Tomcat and Jetty. It turned out that I had bumped into 1 Resin bug (5776) and 2 Tomcat bugs (56736, 56739) related to servlet asynchronous processing support.
Architecturally, Test types 1 and 6 have been implemented using traditional synchronous Servlet API, while the rest of the test implementations leverage non-blocking request handling through Servlet 3 asynchronous processing support. The test implementations store their data in the Apache Cassandra 2 NoSQL database, which is accessed using the DataStax Java Driver. Asynchronous processing is also used in the data access tier in order to minimize resource consumption. JSON data is processed with the Jackson JSON library. In Java versions predating version 8, asynchronous processing requires passing around callbacks in the form of anonymous classes, which can at times be a bit high-ceremony syntactically. Java 8 Lambda expressions does away with some of the ceremonial overhead, but unfortunately TFB doesn’t yet fully support the latest Java version. I’ve previously used the JAX-RS 2 asynchronous processing API, but not the Servlet 3 async API. One thing I noticed during the test implementation was that the mechanism provided by Servlet 3 async API for generating error response to the client is much lower level, less intuitive and more cumbersome than its JAX-RS async counterpart.
The test implementation code was merged in the FrameworkBenchmarks code base, so it should be benchmarked on the next round. The code can be found here:
TechEmpower’s Framework Benchmarks is a really valuable contribution to the web framework developer and user community. It holds great potential for enabling friendly competition between framework developers, as well as, framework users, and thus driving up performance of popular frameworks and adoption of framework performance best practices. As always, there’s room for improvement. Some areas from a framework user and test implementer point of view include: make the benchmark tests and results more repeatable, publish raw benchmark data for analysis purposes and work on making test development and adding new framework components even easier.
Good job TFB team + contributors – can’t wait to see Round 10 benchmark data!