I hope you're outraged that your performance tools are lying to you. For quite a while, many Java sampling profilers have been known to blatantly misrepresent reality. In a nutshell, stack sampling using the documented JVMTI GetStackTrace method produces results that are biased towards safepoints, and not representative of the real CPU processing performed by your program.
Over the years, alternative profilers popped up, trying to fix this problem by using AsyncGetCallTrace, a less-documented API that doesn't wait for a safepoint, and can produce more accurate results. Simply calling AGCT from a timer signal handler gives you a fairly reliable way...
In my last post, I lamented the lack of call stack support for LTTng events in .NET Core. Fortunately, being open source, this is somewhat correctable -- so I set out to produce a quick-and-dirty patch that adds USDT support for CoreCLR's tracing events. This post explores some of the things that then become possible, and will hopefully become available in one form or another in CoreCLR in the future.
Very Brief USDT Primer
USDT (User Statically Defined Tracing) is a lightweight approach for embedding static trace markers into user-space libraries and applications. I've taken a closer look a year ago when...
After exploring the basic profiling story, let's turn to ETW events. On Windows, the CLR is instrumented with a myriad of ETW events, which can be used to tackle very hard problems at runtime. Here are some examples of these events:
Thread start/stop (including thread pool threads)
Exceptions thrown, caught, filtered
Methods compiled (JIT)
By collecting all of, or a subset of, these events, you can get a very nice picture of what your .NET application is doing. By combining these with Windows kernel events for CPU sampling, file accesses, process creations and more -- you have a fairly...
In the same vein of my previous post on analyzing core dumps of .NET Core applications on Linux, let's take a look at what it takes to do some basic performance profiling. When starting out, here are a few things I wrote down that would be nice to do:
CPU profiling (sampling) to see where the CPU bottlenecks are
Grabbing stacks for interesting system events (file accesses, network, forks, etc.)
Tracing memory management activity such as GCs and object allocations
Identifying blocked time and the block and wake-up reasons
With this task list in mind, let's get started!
Collecting Call Stacks of .NET Core Processes
Generally speaking, a...
A lot of high-level languages, runtimes, and libraries people use on Linux have USDT probes embedded in them. In some cases, you have to compile with a specific flag to get the probes embedded (e.g. in Node.js), and in other cases they are part of the default package on most major distributions (e.g. in OpenJDK). Some examples of information these probes can provide include:
Garbage collection events and latencies in Java and Node
Method invocations and latencies in Python and Ruby
Object allocations in Ruby and Java
Thread start and stop events in Java
Class load events in Java and Ruby
BPF is the next Linux tracing superpower, and its potential just keeps growing. The BCC project just merged my latest PR, which introduces USDT probe support in BPF programs. Before we look at the details, here's an example of what it means:
# trace -p $(pidof node) 'u:node:http__server__request "%s %s (from %s:%d)" arg5, arg6, arg3, arg4'
TIME PID COMM FUNC -
04:50:44 22185 node http__server__request GET /foofoo (from ::1:51056)
04:50:46 22185 node http__server__request GET / (from ::1:51056)
Yep, that's Node.js running...
Warning: This post requires a bit of background. I strongly recommend Brendan Gregg's introduction to eBPF and bcc. With that said, the post below describes two new bcc-based tools, which you can use directly without perusing the implementation details.
A few weeks ago, I started experimenting with eBPF. In a nutshell, eBPF (introduced in Linux kernel 3.19 and further improved in 4.x kernels) allows you to attach verifiably-safe programs to arbitrary functions in the kernel or a user process. These little programs, which execute in kernel mode, can collect performance information, trace diagnostic data, and aggregate statistics that are then...
"How much memory is your process using?" -- I bet you were asked that question, or asked it yourself, more times than you can remember. But what do you really mean by memory?
I never thought it would be hard to find a definitive resource for what the various memory usage counters mean for a Windows process. But try it: Google "Windows Task Manager memory columns" and you'll see confusing, conflicting, inconsistent, unclear explanations of what the different metrics represent. If we can't even agree on what "working set" or "commit size" means, how can we ever monitor our Windows...
In my previous post on MiniDumper, I promised to explain in more detail how it figures out which memory ranges are required for .NET heap analysis. This is an interesting story, actually, because I tried a couple of approaches that failed before coming up with the final idea. Basically, I knew that a dump with full memory contains way more information than is necessary for .NET dump analysis. Even if you need the entire .NET heap available, you typically don't need a bunch of other memory ranges: executable code, Win32 heaps, unused regions of thread stacks, and so much more.
Oops! This was sitting in my queue for several months now, and I just noticed it needs to be published. But better late than never I guess. Here goes:
I've been lucky enough to be invited to speak at TechDays Netherlands again this year. This time I was asked to do four talks on some of my favorite subjects -- performance optimization, debugging, and diagnostics. Same as last year, the conference was impeccably organized.
I'm really looking forward to next year's TechDays :-) In the meantime, here are the materials from my talks.
Making .NET Applications Faster
My usual favorite on improving...