Welcome!

Python Authors: Pat Romanski, Matt Davis, AppDynamics Blog, Donald Meyer, Liz McMillan

Related Topics: Java IoT, Microservices Expo, Open Source Cloud, Machine Learning , Release Management , Python

Java IoT: Article

Profiling Python Performance Using lineprof, statprof, and cProfile

Let’s look at profiling in Python

If you’re a regular here, you know how much we care about the full-stack view of applications. Today, let’s zoom in a bit, and talk about the performance of a single layer. In particular, let’s look at profiling in Python.

As an example, let’s take this bit of code for calculating confidence intervals of the mean of a set of data. If you’re not familiar with confidence intervals, they provide a set of bounds for a given statistic; a 95% interval implies that the true mean lies in the calculated range 95% of the time. One way of calculating this involves generating a number of new data sets from the data you have (random selection, with replacement), and looking at that (meta?-)data set. In practice, you typically run this a couple of times, look at the results, and see if they’re converging. If not, you keep running until they do converge, or you hit some max number of iterations.

It turns out, doing 1000 iterations of a reasonable data set size (say, 100k points) is a reasonably expensive operation — about 2 seconds (ish) on my laptop. That’s not too bad in isolation, but run every 15 minutes, on a variety of data sizes up to 1 million, over all customers here at AppNeta? We can do much better.

Since this is Python, let’s see if there’s an obvious way to do it. As it turns out, there is: the standard library module cProfile. This is Python’s flexible, deterministic, C-implemented profiler. (There’s also profile, a pure-python version, but unless you have good cause to avoid C extensions, cProfile is probably a better bet.) cProfile will give us a list of all function calls, sorted however we want. Let’s run it, and dump the result to a file:

ProfilePython-1

There’s a few important things to take note of in these results. The first thing to notice is that 3 of the top 4 functions are in the random module. We only use the random module twice in our example, and the initial use (generating the data set) is a constant 100,000 calls. That leaves 4 million calls from the inner loop of confidence_interval. Since that’s a core part of the algorithm, I wonder if there’s a better way to do that?

For a start, let’s try using a different random number generator. As a rule, if you have a slow function in python, and numpy has a replacement function, the numpy version will be faster. Let’s replace random with numpy.random, and try again:

ProfilePython-2

Much better! Unfortunately, now we have a different problem. Instead of a single line or function call, we only know our problem exists inside a function.

One way to drill down on this is to use statprof. This is a different type of profiler: instead of instrumenting each function call, statprof wakes up at pre-defined intervals and records a stack trace, including line number. Let’s run that:

ProfilePython-3

Cool! Looking back at the code, lines 44 and 45 are the selection and array creation:

ProfilePython-4

Interestingly, the absolute time reported by statprof is a bit lower than cProfile. That difference is due to the statistical nature of statprof. Even at the default setting of 1ms samples, it simply has to do less work than cProfile, who has to instrument all 4 million calls to the various parts of random. It’s worth remembering this difference in overhead, as some code will exihibit different behavior as the timing changes.

In any case, can we do better? It’s possible that the time here is no longer being spent on calculations, but instead on simply creating intermediate arrays. Let’s combine those lines, and re-run:

ProfilePython-5

That certainly seems better: 1.92s vs 2.39s (1.54 + 0.85) from above. At this point, though, be careful. Up until now, we’ve avoided directly comparing profiling runs. Looking at the percentage of time spent, we’re actually at about the same place (85% of total); this run just completed faster than the last run. This is a common pattern — as the low-hanging fruit is picked off, it becomes more and more important to validate smaller gains in more formal ways. How, exactly, to collect those stats is a topic for another post.

It’s also worth mentioning Robert Kern’s line_profiler. It’s a deterministic profiler (like cProfile), but it captures profiling data on a line-by-line basis. It also allows selectively profiling specific functions, which can be invaluable on a larger codebase, to weed out the noise from other modules. Unfortunately, because it does capture every line call you ask it to, the overhead of instrumenting every line you run can be substantial (4x slowdown on my laptop). For completeness, here’s what it looks like on our sample:

ProfilePython-6

Finally, remember that profiling is just one component of a performance toolkit. Most performance problems aren’t complex; they’re just well-hidden. Before writing the above code, I hadn’t realized how expensive random number generation could be. Simply taking the time to look for a faster function saved over 90% of the time spent.

Make it work, then make it pretty, then make it fast. And with the right tools, it doesn’t even take that long.

And if you want to find other ways to optimize your code and make it fast, start using TraceView for free - sign up here.

Related Articles

Python and gevent

Tracing Celery Performance For Web Applications

Tracing Python — An API

More Stories By TR Jordan

A veteran of MIT’s Lincoln Labs, TR is a reformed physicist and full-stack hacker – for some limited definition of full stack. After a few years as Software Development Lead with Thermopylae Science and Techology, he left to join Tracelytics as its first engineer. Following Tracelytics merger with AppNeta, TR was tapped to run all of its developer and market evangelism efforts. TR still harbors a not-so-secret love for Matlab-esque graphs and half-baked statistics, as well as elegant and highly-performant code. Read more of his articles at www.appneta.com/blog or visit www.appneta.com.

IoT & Smart Cities Stories
IoT is rapidly becoming mainstream as more and more investments are made into the platforms and technology. As this movement continues to expand and gain momentum it creates a massive wall of noise that can be difficult to sift through. Unfortunately, this inevitably makes IoT less approachable for people to get started with and can hamper efforts to integrate this key technology into your own portfolio. There are so many connected products already in place today with many hundreds more on the h...
CloudEXPO New York 2018, colocated with DXWorldEXPO New York 2018 will be held November 11-13, 2018, in New York City and will bring together Cloud Computing, FinTech and Blockchain, Digital Transformation, Big Data, Internet of Things, DevOps, AI, Machine Learning and WebRTC to one location.
SYS-CON Events announced today that IoT Global Network has been named “Media Sponsor” of SYS-CON's @ThingsExpo, which will take place on June 6–8, 2017, at the Javits Center in New York City, NY. The IoT Global Network is a platform where you can connect with industry experts and network across the IoT community to build the successful IoT business of the future.
Andrew Keys is Co-Founder of ConsenSys Enterprise. He comes to ConsenSys Enterprise with capital markets, technology and entrepreneurial experience. Previously, he worked for UBS investment bank in equities analysis. Later, he was responsible for the creation and distribution of life settlement products to hedge funds and investment banks. After, he co-founded a revenue cycle management company where he learned about Bitcoin and eventually Ethereal. Andrew's role at ConsenSys Enterprise is a mul...
DXWorldEXPO | CloudEXPO are the world's most influential, independent events where Cloud Computing was coined and where technology buyers and vendors meet to experience and discuss the big picture of Digital Transformation and all of the strategies, tactics, and tools they need to realize their goals. Sponsors of DXWorldEXPO | CloudEXPO benefit from unmatched branding, profile building and lead generation opportunities.
Disruption, Innovation, Artificial Intelligence and Machine Learning, Leadership and Management hear these words all day every day... lofty goals but how do we make it real? Add to that, that simply put, people don't like change. But what if we could implement and utilize these enterprise tools in a fast and "Non-Disruptive" way, enabling us to glean insights about our business, identify and reduce exposure, risk and liability, and secure business continuity?
The best way to leverage your Cloud Expo presence as a sponsor and exhibitor is to plan your news announcements around our events. The press covering Cloud Expo and @ThingsExpo will have access to these releases and will amplify your news announcements. More than two dozen Cloud companies either set deals at our shows or have announced their mergers and acquisitions at Cloud Expo. Product announcements during our show provide your company with the most reach through our targeted audiences.
DXWorldEXPO LLC announced today that Telecom Reseller has been named "Media Sponsor" of CloudEXPO | DXWorldEXPO 2018 New York, which will take place on November 11-13, 2018 in New York City, NY. Telecom Reseller reports on Unified Communications, UCaaS, BPaaS for enterprise and SMBs. They report extensively on both customer premises based solutions such as IP-PBX as well as cloud based and hosted platforms.
To Really Work for Enterprises, MultiCloud Adoption Requires Far Better and Inclusive Cloud Monitoring and Cost Management … But How? Overwhelmingly, even as enterprises have adopted cloud computing and are expanding to multi-cloud computing, IT leaders remain concerned about how to monitor, manage and control costs across hybrid and multi-cloud deployments. It’s clear that traditional IT monitoring and management approaches, designed after all for on-premises data centers, are falling short in ...
The deluge of IoT sensor data collected from connected devices and the powerful AI required to make that data actionable are giving rise to a hybrid ecosystem in which cloud, on-prem and edge processes become interweaved. Attendees will learn how emerging composable infrastructure solutions deliver the adaptive architecture needed to manage this new data reality. Machine learning algorithms can better anticipate data storms and automate resources to support surges, including fully scalable GPU-c...