Thursday, June 9, 2022

How Many Eggs Can You Fit In One Nest?

Prof. Thain gave a talk at HTCondor Week 2022, giving an overview of some of our recent work on resource management in high throughput scientific workflows.  An HTCondor talk requires a "bird" metaphor, so I proposed the following question:

How many eggs can you fit in one nest?



A modern cluster is composed of large machines that may have hundreds of cores each, along with memory, disk, and perhaps other co-processors.  While it is possible to write a single application to use the entire node, it is more common to pack multiple applications into a single node, so as to maximize the overall throughput of the system.

We design and build frameworks like Work Queue that allow end users to construct high throughput workflows consisting of large numbers of tasks:

But, how does the end user (or the system) figure out what resources are needed by each task?  The end user might have some guess at the cores and memory needed by a single task, but these values can change dramatically when the parameters of the application are changed.  Here is an example of a computational chemistry application that shows highly variable resource consumption:

CCL grad student Thanh Son Phung came up with a technique that dynamically divides the tasks into "small" and "large" allocation buckets, allowing us to automatically allocate memory and pack tasks without any input or assistance from the user:

Here is a different approach that we use in a high energy physics data analysis application, in which a dataset can be split up into tasks of variable size.  Instead of taking the tasks as they are, we can resize them dynamically in order to achieve a specific resource consumption:

Ben Tovar, a research software engineer in the CCL, devised a technique for modelling the expected resource consumption of each task, and then dynamically adjusting the task size in order to hit a resource target:

To learn more, read some of our research research papers:

Monday, February 7, 2022

IPDPS Paper: Dynamic Task Shaping ... in High Energy Physics

In an upcoming paper to be presented at IPDPS 2022, we discuss our experience with designing and executing high throughput data intensive applications for high energy physics.  The application itself is pretty cool: TopEFT is a physics application that uses the Coffea framework for parallelization, the Work Queue framework for distributed execution, and XRootD for remote data access:


Configuring such applications to run on large clusters is a substantial end-user problem.  It's not enough to write a correct application: one must also select a wide variety of performance parameters, like the data chunk size, the task length, the amount of memory per task, and so on.  When these are chosen well, everything runs smoothly.  But even one parameter out of tune can result in the application taking orders of magnitude longer than necessary, wasting thousands of resources resources, or simply not running at all.  Here is the end-to-end runtime for a few configurations with slight variations:

This paper describes the techniques that we have developed to make TopEFT a self-configuration application: dynamic task sizing, resource monitoring, modeling of resource consumption, and more.   A particular challenge is that the tasks are not identical, but they do exhibit some rough linearity:

With these techniques, we are able to relieve the user of the burden of setting a variety of controls, and allow the system to find its own stable configuration.  Check it out:


Scaling Up Julia: Hidden Filesystem Stress

HTCondor Cluster View
In the CCL, we study the design and implementation of scalable systems and applications that run on very large computing systems.  It is not unusual for us to encounter an application that runs well on a few nodes, but causes trouble when running on thousands of nodes.  This happened recently with a simulation written in Julia that was using HTCondor to run millions of tasks running on several thousands nodes on our campus cluster.  It ran fine on one node, but when deployed to thousand nodes, this simulation would cause a total meltdown of the shared filesystem, even though its I/O needs were relatively small.  What was going on?

Here is what we found:

The Julia programming language uses a just-in-time compiler to generate efficient machine code before execution. Julia organizes code in modules, and user applications in projects, where a project is a list of modules. By default, the compilation step is performed every single time an application is executed and considers all the modules listed in the given project.  If an end user sets up an application in the normal way, the result is that the code will be compiled simultaneously on all nodes of the system!

Internally, Julia checks the project's list of modules, checks for files with a modification time more recent than the machine code already available, and if needed, generates new machine code.  As usual, the modifications times are done using the stat() system call. To give some perspective, the simulation used a dozen standard Julia modules, resulting in 12,000 stat() calls  even when no recompilation was needed. But the number of open() calls to needed files was less than 10. In particular, the file that listed the modules in the project (Project.toml) had close to 2,000 stat() calls, but only one open() call. For comparison, the number of calls to open() and stat() for data files particular to the application was less than 5.

When executed in a single machine on a local file system, even a few thousand system calls may unnoticed by the user. However, they become a big problem when trying to run at scale in a cluster where all nodes share a common networked filesystem. If one thousand nodes start at once, the shared filesystem must field twelve million stat() operations just to determine that nothing has changed.  Thus, the scale at which the simulation can run will be limited by factors hidden to the end user, that is, not by the cores, memory, or disk available, but by these file system operations that become expensive when moving from a local to a shared setting.

Once the problem is understood, the workaround is to pre-compile a binary image with the needed modules that then is shipped together with each task.  This reduced the number of stat() calls from the original 12,000 to about 200 per invocation.  This is image is shipped compressed with each job, to reduce its size from 250MB to 50MB, and decompressed just before the task start execution.  Generating the binary image takes about 5 minutes, prior to job submission.

The user application made the generation of the binary image much easier because all the dependencies were listed in a single file. As an example, consider this file that simply lists some modules:

# my_modules.jl
using Pkg
using Random
using Distributions
using DataFrames
using DataStructures
using StatsBase
using LinearAlgebra
If we count the number system calls that involve filenames, we get:

$ strace -f -e trace=%%file julia my_modules.jl |& grep  -E '(stat|open)'| wc -l
5106


These calls will be repeated everytime the program runs. Using the module PackageCompiler we can generate a Julia system image as follows:

# comp.jl

# run as: julia comp.jl
loaded_by_julia = filter((x) -> typeof(eval(x)) <:  Module && x ≠ :Main, names(Main,imported=true));

include("my_modules.jl")

loaded_all = filter((x) -> typeof(eval(x)) <:  Module && x ≠ :Main, names(Main,imported=true));
loaded_by_ch = setdiff(loaded_all, loaded_by_julia);

println("Creating system image with:");
println(loaded_by_ch);

using PackageCompiler;
create_sysimage(loaded_by_ch; sysimage_path="sysimage.so", cpu_target="generic")

Using the image, the number of file releated calls, and there the stress on the
share file system,  are greatly reduced:

$ strace -f -e trace=%file julia -Jsysimage my_modules.jl |& grep -E '(stat|open)' | wc -l
353


Also, as expected, the overhead per run also decreases, as the runtime decreases from about 10s to about 0.5s, which is significant for short running tasks.

So what's the moral of the story?

1 - When moving from a single node to a distributed system, operations that were previously cheap may become more expensive.  You can't fix what you can't measure, so use tools like strace to understand the system-call impact of your application.

2 - Avoid exponential behavior, even when individual costs are cheap.  Every Julia import results in checking the freshness of that module, and then all of its dependencies recursively, and so leaf modules get visited over and over again.  The Julia compiler needs to memoize those visits!