Tools for live extrospection of the Erlang BEAM VM — WARNING: early alpha
Switch branches/tags
Clone or download
tokenrove Explicitly link libdw
This solves a linking error in the Gentoo ebuild.  It's an unfortunate
hack that is thanks to our using our own libelf, which we should stop
doing.
Latest commit 6f84bf3 Feb 2, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
src
vendor
.gitignore
.gitmodules
README.md
build.ninja
build.sh
meson.build
meson_options.txt

README.md

think-outside-the-beam

This is a collection of tools for unobtrusive introspection of a running Erlang VM under Linux. It uses Linux-specific interfaces (perf events, process_vm_readv(2)) to avoid having to stop the process with ptrace(2).

Because these tools are necessarily approximate (see the WARNING section), they should be used as a way to discover new directions for more specific investigation, not as a source of truth.

WARNING

These tools make significant assumptions about the internals of the Erlang VM they are introspecting. They probably won't work on even slightly different versions of the VM. They are also very specific to x86-64 and Linux presently.

Compatible version: Erlang OTP 18.3 x86-64

These tools are also necessarily inaccurate. First, perf itself may sample in a biased fashion; secondly, these tools read additional data from the VM, and may receive inconsistent or garbled views of the data therein. Use them only for developing hypotheses.

EXTRA WARNING

This software has only been used on a handful of fairly homogeneous systems, by the author. It is alpha software that is almost certainly broken in many subtle ways.

To build

Try:

./build.sh

You will need:

We will eventually supply a script that verifies the constants chosen here are consistent with the running BEAM internals.

You will have the best luck if you've built your Erlang system with at least these flags:

-fvar-tracking-assignments -ggdb -g3 -gdwarf-4 -Wl,--build-id

Standalone Tools

Most of these tools by default look at all threads associated with a VM, but can be asked to isolate only a single PID.

Those that have some amount of skid (accuracy of measurement) can be asked to print estimates of how far off they're likely to be (--skid-summary).

At some point, generating the perf.map will be done automatically, but for the moment, it must be done manually. So before running erlang-sample, you'll need to run erlang-write-perf-map PID where PID is the PID of your Erlang VM. If it fails because of missing symbols, you'll probably need to rebuild Erlang with debugging options.

erlang-sample

By default, lists most frequently seen Erlang function calls as sampled from a running VM.

--pstack

Prints the stack of each running process on each scheduler. By default, only those schedulers that are running. Options to wait for each scheduler; print only Erlang stack.

--blame

For a given native function (like erts_garbage_collect or copy_struct), report those Erlang functions occurring most frequently in the stack trace for that function.

erlang-heapsample (coming soon)

Print heap information about currently running processes in Erlang VM.

Integrations

perf

See vendor/perf. Still extremely nascent.

kcov

Coming soon, hopefully.

Questions and How to Answer Them (WIP)

What are the hottest Erlang functions?

Run erlang-sample for a reasonable period of time, probably with the --only-erlang option.

What is allocating long-lived memory?

How much are NIFs impacting scheduling?

Where do expensive deep copies occur?

Try erlang-sample --blame copy_struct.

(when copy_struct is seen, dump the stack of the process involved)

How much RSS/vsize is lost to fragmentation?

  • compare maps and mbcs carriers, sbcs carriers to actual memory allocated

There may be some metric we can come up with for this, too.

(total free - largest free block) / total free

See also recon_alloc:fragmentation for the in-VM approach to this.

We should also be able to compare memory usage to actual vsize of anonymous rw pages.

How is my workload distributed across schedulers? Across CPUs?

Graph processes seen and percentages of other stuff, per scheduler

How this works

Erlang stacktraces using perf event sampling and process_vm_readv

There are two mechanisms by which we get information from the VM process: perf event sampling, which is done by the kernel synchronously (as far as I know), and direct reading of the VM process's memory using process_vm_readv, which necessarily happens asynchronously and can present an inconsistent picture of the VM's state.

(See "Why not ptrace or /proc/PID/mem?" elsewhere in this documentation, if you just asked yourself that question.)

We're mostly concerned with getting samples when the native IP is inside process_main, although it doesn't hurt to get the most accurate backtrace possible even if we're in some child of process_main like erts_garbage_collect.

In process_main, we have a couple of variables that are particularly of interest. There's c_p, which points to the current process. We can read all kinds of useful information from that structure, but (except with some dirty tricks that aren't generally applicable) the time between a perf sample being made and us reading this information could be very large (see other discussions in this documentation on skid about that).

So, if we can, we also want to sample I and E. I points to the current instruction, and E points to the top of the stack. If we can get all of them, we can do a pretty good job of validating that the trace we read from c_p is accurate with regards the perf sample.

Tell me about the dirty tricks that aren't generally applicable

This is probably one hack too far, but consider if we get perf to sample the stack of the following bit of code:

    spy_pid = syscall(__NR_gettid);
    sched_setscheduler(spy_pid, SCHED_IDLE,
                       &(struct sched_param){.sched_priority=20});
    asm volatile("" ::: "memory");
    /* this should probably be nanosleep, but since we destroyed our
     * stack forever, we'd have to put the arguments in static storage
     * or similar.  too much hassle for this prototype.  sched_yield
     * shouldn't be _so_ bad if there are other jobs to run. */
    asm volatile ("forever:\n"
                  "movq %0, %%rsp\n"
                  "movl %1, %%eax\n"
                  "int $0x80\n"
                  "jmp forever\n"
                  : : "r" (spy_target), "r" (__NR_sched_yield) : "rsp");
    __builtin_unreachable();

This allows us to sample memory from wherever we point spy_target. (For example, we could write a NIF that allows us to create these spy threads in the VM and then read and write their spy_target with process_vm_{readv,writev}.) So we might be able to use this to sample a single process with a higher level of accuracy than before, but it would require some serious juggling and machinations that don't seem to be worth it.

At this point, if you're considering doing this, you probably just want to extend perf's sampling mechanism in the kernel. SystemTap or the new BPF facilities probably are better places to aim for this.

Why not ptrace or /proc/PID/mem?

It's fairly well-known that in order to ptrace, we have to stop the traced process. (Disclaimer of ignorance: I know that PTRACE_SEIZE exists as a Linuxism, but I don't know how much you can do in that state without invoking PTRACE_INTERRUPT.)

It's less well-known that in order to read from /proc/PID/mem, the same is true: the process must be stopped.

Most of the systems I was interested in applying these techniques to cannot abide being stopped even briefly.

Troubleshooting

In general, the --pstack mode for erlang-sample is useful for troubleshooting, since it prints full stack traces at a time and one can easily see many common problems (such as all traces being a single entry deep, or no Erlang functions ever appearing).

erlang-sample can't find a register location for c_p

If the problem is just that the register information is more complex than a single location (you can check with dwarfdump, readelf or similar), it's mostly a matter of making erlang-sample smarter.

If the location isn't there at all, though, (i.e. gdb gives the dreaded (optimized out) message when you do info address c_p when stopped in process_main) you can find this and other important registers by looking at the disassembly of process_main.

For example, one can run objdump -d -S /usr/local/lib/erlang/erts-7.3.1/beam.smp | less (replace with a suitable path to your copy of beam.smp or beam), search for process_main, then within process_main, look for the disassembly immediately following macros like SWAPIN. Chances are, you'll see something like this:

        SWAPIN;
  43e01c:       4d 8b 55 50             mov    0x50(%r13),%r10
  43e020:       ff 23                   jmpq   *(%rbx)
  43e022:       49 8d 95 c8 02 00 00    lea    0x2c8(%r13),%rdx

From that, it seems pretty likely that r13 is c_p, and rbx is I. We could be wrong, of course, but we can test it out with:

erlang-sample --force-c_p-register=r13 --force-I-register=rbx --pstack -d 1 PID

and see if the results are at all sensible.

erlang-sample doesn't seem to be able to unwind (no backtraces)

Unfortunately at the moment we rely on our slighly-hacked vendored copy of elfutils, which causes as many problems as it solves. Try

LD_LIBRARY_PATH=vendor/elfutils/backends ./build/dist/erlang-sample --pstack PID

and see if it's any better. elfutils may not be able to find the suitable EBL backend, which it always loads dynamically even if libdw was statically linked into the program.

Open Problems

How much skid is there in a given measurement, and how can we reduce it?

See skid measurement options.

When we receive an actual process_main sample, or something where that's in the call stack, we actually have more information than it might seem.

We can sample E and I when they're in registers. There's a bunch of other corroborating evidence. For example, we can look at what opcode we were executing in process_main, and try to correlate it with opcodes in the source of the processes on that scheduler.

Can we avoid depending on -ggdb builds by writing the perf map from the VM itself?

We still need to know where c_p, E, I, and so on live, which requires either DWARF or manual inspection of the source (or perhaps some automated reverse engineering).