Series Introduction: Overload and how Wallaroo mitigates overload
This is the first of a pair of Wallaroo Labs articles about overload. Here’s a sketch of the series.
- Part one presents a brief overview of queueing networks and what the term overload means for a queueing network, followed by an outline of some common techniques that computer systems use to mitigate the effects of overload.
- Part two discusses the details of how Wallaroo uses one of those techniques, back-pressure, to manage overload conditions.
Part two of the series is now available at: How the end-to-end back-pressure mechanism inside Wallaroo works.
Introduction: An overview to queueing networks
Most of the details of the mathematics of queueing theory aren’t needed for in this article or for the follow-up article. However, a very small introduction to queueing theory will help put us all on the same page.
Here’s a diagram of a very simple queueing system.
This basic model has been applied to many different types of systems, including:
- Phone calls in telephone exchanges
- Waiting at the deli counter or at the post office
- Street and expressway traffic congestion
- Computer systems, especially for capacity planning purposes
In the case of a deli counter or the post office, you might know (or
assume) that deli/postal service requires 60 seconds per customer, on
average. If customers arrive at a rate of
X people per minute, then
how long will the queue time typically be, you wonder? Or, perhaps
you want to predict how many people will usually be waiting in the queue.
Mathematicians have worked for roughly 100 years to define scenarios & models for many different kinds of arrival rate schemes, service time schemes, and queue size limits. It’s surprising how powerful a tool this single abstraction is, one queue plus one service. Furthermore, when the model’s simplicity is found lacking, mathematicians started putting multiple queue+service centers together into networks to solve bigger problems. Here’s a multi-service center network model:
Here is a network model for the edit-compile-execute cycle of software development:
I’m intentionally omitting all of the many methods that are used to predict answers to questions like, “How long will the queue be?” or “How long will I wait?” Instead, I wish to highlight one of the fundamental assumptions that most of those methods require: steady state.
The assumption of steady state might instead be called “flow balance assumption” or “stability”, depending on the book or paper that you’re reading. A steady state network queue model has an arrival rate less than or equal to the departure rate. If a system is not in steady state, then one (or more!) queues in the model’s network will grow without limit.
What does a steady state assumption mean for your system? From a theoretical point of view, a violation of steady state means that the model loses its predictive power: the model cannot tell you how the system will behave. From a practical point of view, an overloaded system has queue sizes that grow without stopping.
We do not yet have computers with truly infinite memory. Until that day, let’s look at how to design systems that try to remain in steady state.
Overload: How to define it and how to mitigate its effects
Let’s use an informal definition for the word overload.
When a finite size service queue becomes full, or when a service queue’s size continues to grow without stopping, then the service is overloaded.
The cause of overload boils down to a simple equation. Note that it is the opposite of the definition of steady state.
Arrival Rate > Departure Rate
If this simple equation is true for short periods of time, then a system is not necessarily overloaded. Is your local post office overloaded if the arrival rate exceeds the departure rate for a 1 minute time period? Usually, no. However, what if 200 customers arrived in that one minute? Then perhaps you declare the post office overloaded if the queue cannot fit inside of the building and spills out onto the sidewalk.
Let’s look at some methods for mitigating overload conditions. We’ll see that the methods available depend on changing the conditions of the steady state equation or on increasing the system’s storage space.
Solution 1: Add more queue space
If an service’s queue gets full, then we simply add more space. Most modern operating systems will do this for you, via the wonderful magic of virtual memory.
We know that computer systems do not have infinite storage
space. If the
Arrival Rate > Departure Rate equation is true for
long enough, any single machine will run out of space. The same
statement holds true for a multi-machine system.
If we add more space without changing the balance of the steady state
equation, then we are simply delaying when the consequences of full
queues will strike us.
You will probably ought to consider an alternative technique.
Solution 2: Increase the Departure Rate
Two common strategies to increase
Departure Rate are
increasing service throughput or decreasing service latency.
Strategies are commonly used to implement them are horizontal scaling
and load shedding, respectively.
Increase the Departure Rate by using faster machines
I’ll mention this option briefly. Using faster computers (and/or
storage devices, networks, etc.) is frequently an excellent way to
Departure Rate of your service. If your service is
already running on slower hardware, then you also have a migration
problem to solve. That migration problem pops up in other
techniques discussed here, so let’s move on to the next technique.
Increase the Departure Rate by horizontal scaling (a.k.a. make your cluster bigger!)
Sometimes, it is possible to add more space (RAM, NVRAM, disk, etc.) or CPU capacity to a single machine. In many of today’s data centers, it is far easier to add additional machines. Or add virtual machines (VMs). Or add containers.
Azure, Google, Amazon, and many other service providers have APIs that include a “Add More, Just Click Here!” feature. (But perhaps not with that exact name.) Adding more capacity to your system is fantastic for the future, but it cannot help now. The extra capacity cannot help your overloaded system right now, because:
- Adding extra capacity may be impossible. For example, the API is easy to use, but the data center is full, which causes the API requests fail.
- You may wait a long time before extra capacity is available.
- Adding extra capacity probably costs more money.
- Overhead from adding capacity may reduce the capacity of the existing system during the transition time.
To be effective, you need to plan ahead. You have to choose some earlier time to start the process of adding capacity, before it’s too late to be helpful. And also, be careful not to act too hastily and/or to add too much capacity. It’s not an easy balance to find and maintain.
Increase the Departure Rate by load shedding
Load shedding is another way to increase the
Departure Rate side of our
service equation. Load shedding implementations can include:
- Do not compute the requested value, but instead send an immediate reply to the client. Usually this reply that also signals that the system is overloaded. (Will clients act upon the overload signal and actually change their behavior? Good question.)
- Choose an alternative computation that requires less time.
- If the service is text search, then only search 10% of the text corpus instead of the full 100%.
- If the service calculates square roots with 25 digits of precision, then reduce the precision to 5 digits instead.
- Drop the request (or query or packet). Do nothing more, literally, as quickly as possible.
Solution 3: Decrease the Arrival Rate
It’s unfortunate, but many computer systems have very little control over a service’s arrival rate. Perhaps a flash mob of 200 customers arrives in one minute at your local post office? You don’t have full control over your customers and their arrival rates, but you likely still have options.
Decrease the Arrival Rate by filtering out some requests
Earlier this month, on
March 1st, 2018, GitHub experienced a record setting denial-of-service attack.
At its peak, the attack generated 1.35 terabits/second of network
traffic at 126.9 million packets/second.
The incident report from GitHub’s Engineering department
explains how Akamai’s services were used to
Arrival Rate by filtering out millions of packets per
second of junk.
GitHub’s systems and Akamai’s systems cooperated to keep GitHub’s data
services usable by GitHub customers.
From GitHub’s point of view, the
Arrival Rate was reduced by Akamai’s
filtering of the workload before it arrived at GitHub’s servers.
From Akamai’s point of view, Akamai acted as load
shedding system; until the attackers relented and stopped their
Arrival Rate remained record-breakingingly high.
Decrease the Arrival Rate by back-pressure (“Hey, customers, stop!”)
Many decades of computer systems research have given us a lot of rate limiting schemes. Most are based on an idea of fake money or credit or tokens or a ticket that a customer must have before the customer can be admitted to a queue. Without the credit/money/token/ticket, then the customer isn’t permitted into the system. Admission control and flow control are two common names for these schemes.
I’m guessing that most of my audience knows a little bit about
the TCP protocol.
TCP includes two mechanisms for controlling
One is TCP’s sliding window protocol, which permits a limited
number of network packets to be in transit in the network without
overloading the receiving system’s capacity.
When the sliding window is non-zero, the sender is permitted to send
some bytes to the receiver, up to the window’s size (in bytes).
When the sliding window is zero, the sender must stop sending.
I’ll have a much more detailed example of TCP’s sliding window protocol in next week’s follow-up article. My apologies, please hold on for part two!
Decrease the Arrival Rate by … doing nothing? (Ride out the storm)
If your queue sizes are large enough, and if
Arrival Rate > Departure Rate is true only for a short amount of
time, then perhaps you can simply do nothing. Simply wait
Arrival Rate to drop.
Perhaps your system is busiest after suppertime, and
Arrival Rate naturally drops when your customers start going to sleep in the
evening. (If your customers tend to eat and sleep at similar times!)
If you can predict your customer’s peak
Arrival Rate with 100%
accuracy, congratulations, you live in a wonderful world.
Otherwise, you probably ought to consider an alternate technique.
Conclusion of part one: Wallaroo’s choices of overload mitigation techniques
Wallaroo Labs has chosen back-pressure as the primary overload mitigation technique for Wallaroo applications. TCP’s sliding window protocol is part of Wallaroo’s end-to-end back-pressure system. When a TCP connection’s advertised window is zero, the receiver is telling the sender, “I am overloaded. You must stop sending now. I will tell you when you can send more.”
The load shedding technique is a poor fit for Wallaroo’s goal of
accurately processing 100% of the data stream(s) without loss.
Wallaroo systems can grow and shrink horizontally today,
but full integration with data center or cloud services management
APIs is a future feature, not available yet.
We believe that waiting for
Arrival Rate to drop on its own isn’t
a good idea.
Next week’s follow-up to this article will detail how Wallaroo’s back-pressure mechanisms work together with TCP’s sliding window protocol to mitigate overload conditions. Thanks for reading! I hope you’ll read Part Two next week.
More material on how to deal with overload
Here are some articles and presentations that you might find useful places to learn more. If you were to read only two items to learn more about handling overload, my recommendations are:
Fred Hebert’s blog, especially Fred Hebert’s “Queues Don’t Fix Overload” which is linked below. I love his illustrations with kitchen sinks filling & draining with water.
I call it the “SEDA paper”, but its proper title is “Adaptive Overload Control for Busy Internet Servers” by Welsh & Culler. It is also linked below. This paper from 2003 is one that I believe everyone ought to read; its ideas will color your thoughts on software design for many years to come.
Here’s the full list. Please explore!
- dataArtisans: How Apache Flink™ handles backpressure
- Fred Hebert: Queues Don’t Fix Overload. If you like this article, Fred’s follow-up article is called is Handling Overload,
- Henn Idan: Reactive Streams and the Weird Case of Back Pressure
- Reactive Streams initiative: Introduction to JDK9 java.util.concurrent.Flow
- Zach Tellman: Everything Will Flow,
an overview of Clojure’s
- Matt Welsh and David Culler: “Adaptive Overload Control for Busy Internet Servers”
- Wikipedia: Admission Control, Back-Pressure, Queueing theory, and Sliding window protocol topics
The queue network figures in this article are excerpts from the book “Quantitative System Performance” by Lazowska, Jahorjan, Graham, and Sevcik, Prentice-Hall, Inc., 1984. Full text of this book is available online.
Did you enjoy this post?
Subscribe to our blog and be updated each time we add a new post.