Microservices design is a key component of modern app development. This architecture, which involves composing multiple (mostly) HTTP or gRPC services, offers numerous advantages, including code reuse, maintainability, scalability, and fault tolerance.

In the context of this post, scalability refers to the software's ability to optimize performance based on available resources, while fault tolerance ensures that the system remains operational even when certain components fail. The interplay between these two aspects becomes crucial when dealing with microservices.

Little’s Law: Decoding Server Performance

Little’s Law is a foundational theorem in operations research and system analysis that offers insights into determining the capacity of stable systems that manage and process external requests. This theorem elegantly connects three pivotal variables:

• λ represents the average rate of incoming requests.
• W denotes the average time taken to process each request.
• L is the number of concurrent requests the system handles.

To illustrate, consider a system that receives 500 requests per second, with each request taking 0.5 seconds to complete. This means the system is concurrently handling 500 × 0.5 = 250 requests.

In this scenario, L represents the system’s capacity, while W is determined by the software’s design, its complexity, and inherent latencies. By knowing both L and W, we can calculate the maximum request rate the system can sustain:

To handle more requests, we need to increase L, the system’s capacity, or decrease W, the processing time or latency. However, if the incoming request rate surpasses λ, the system will no longer be stable. Requests will start queuing up, initially experiencing much-increased latency, but quickly leading to full consumption of system resources and service unavailability.

Understanding Capacity (L)

The capacity, L, of a web server is not a constant value. It’s variable, influenced by several environmental (hardware, operating system (OS), etc.) and limiting factors:

• TCP Connections: Typical servers can handle tens of thousands of simultaneous TCP connections, with the possibility of sustaining millions of active connections in certain cases with the right configuration.
• Bandwidth: Except for streaming HD video, most requests and responses over LAN are just a few kilobytes in length. Given that the total volume of a single request is under 1MB, a network can support anywhere from 100K to over a million concurrent requests.
• RAM: Assuming each request uses less than 1MB of memory, and considering the affordability of RAM, it’s feasible to support over a million concurrent requests, or at the very least, several hundreds of thousands.
• CPU: Given that most of the processing is handled by microservices and requests primarily wait for responses, CPUs can efficiently manage anywhere from several hundreds of thousands to multiple millions of concurrent requests. In real-world systems with similar architectures, CPU usage is seldom the limiting factor.

From our analysis of the limiting factors, we expect to sustain L in the range of 100K to about 1 million. However, the operating system (OS) introduces a key constraint. When using the thread-per-request model, each request is processed on a single OS thread from start to finish, meaning our ability to manage concurrent requests is directly limited by the number of threads the OS can handle.

In the thread-per-request model, threads perform some processing, then block while waiting for a microservice or other downstream services to respond, and repeat this process. They’re not constantly active – not fully utilizing the CPU – but they aren’t entirely idle either. The OS typically schedules each thread multiple times for a single request. During these idle times, the OS temporarily stops the waiting thread and switches to another task, then reschedules it once a response is received. Given this behavior, most OS typically supports between 2K and 20K concurrent threads. Going beyond this limit introduces latency, causing W to rise and λ to fall. To prevent system strain from excessive thread creation, there’s often a cap on the number of threads that can be spawned, usually set between 500 and 20K, but rarely more than that.

Because L is the minimum of all these limits, the OS scheduler suddenly dropped our capacity, L, from the high 100Ks to ­low millions, to well under 20,000.

If we use the thread-per-request model on a “good-­enough” hardware, L is largely determined by the OS’s thread capacity without adding latency. While purchasing additional servers is an option, it’s a costly one with potential hidden cost. Hesitation increases when we realize that software is the problem, and our existing servers remain underutilized.

Processing Latency (W)

Latency is a crucial factor in determining system efficiency. Imagine two microservices, Service A and Service B, each taking an average of 500ms to respond, inclusive of network latency. If the services are being called sequentially, the web service’s processing latency is 1 second, denoted as W. If we permit the web server to generate up to 2000 threads, our L becomes 2000. By Little’s law, our system can manage:

requests per second before instability ensues. This seems sufficient, even accounting for traffic surges.

However, this holds true only when both Service A and Service B function optimally. If one’s latency spikes to 10 seconds, W shifts from 1 second to approximately 10 seconds (From W = 0.5 + 0.5 = 1 we’ve now gone to W = 0.5 + 10 = 10.5). Consequently, λ drops from 2000 to 200 requests per second, which is suboptimal.

To make our service fault-tolerant, let’s set timeouts for the service calls — say we set our HTTP client’s timeout to 2 seconds. This ensures our maximum latency, even amidst failures, remains at 4 seconds, resulting in a λ of 2000/4 = 500 requests per second.

We still have room for improvement. If Service A consistently fails and times out, it’s inefficient to keep trying to connect to it, especially with a 2-second wait for each attempt. Instead, we can implement a “circuit breaker.” When Service A fails, the circuit breaker activates, halting further attempts to connect. Periodically, a background task checks Service A’s status. If it’s operational, the circuit breaker deactivates. This strategy can reduce our latency to less than 1 second and maintain our capacity at 2000 requests per second, even if failures occur.

Additionally, some circuit breaker libraries allow for allocating specific thread pools for different tasks, ensuring a single malfunctioning operation doesn’t consume all available threads. This approach is known as “bulkheading” against failures.

Can we further reduce W? Yes, if Service A’s result isn’t a prerequisite for Service B, we can parallelize their calls. Using Futures (Java) or Promises (JS), we dispatch both requests simultaneously, potentially reducing W from 1 second to 500ms. This principle extends to multiple service calls, allowing our server to manage up to 4000 requests per second, even with numerous microservice interactions.

Is this our peak performance? Barring significant optimizations to Service A and Service B, it appears so. Even with underutilized hardware, 4000 requests per second is our upper limit, especially if any microservice latency exceeds 500ms. If this isn’t sufficient, what’s our next move?

Optimizing Latency with Functional Callbacks

While we’ve managed to reduce W, L remains limited by the OS’s thread-handling capacity. To increase L, we must move away from the thread-per-request model.

Enter Node.js, a server-side JavaScript framework. Given JavaScript’s single-threaded nature, Node.js bypasses the OS thread scheduler. Instead of allocating a thread for each request, it uses asynchronous callbacks. When a request handler needs to wait (e.g., for a service call), it provides a callback to execute upon completion. This approach capitalizes on JavaScript’s lack of threading, avoiding the OS thread limitations.

However, this method has its challenges. Any prolonged blocking of a handler function stalls the entire Node.js instance, and no other requests can be processed. This is often mitigated by running several Node instances on one machine and load-­balancing them, which also helps take advantage of all CPU cores. However, this consumes more RAM and makes parallelizing certain computational tasks difficult. Moreover, the asynchronous, callback-based programming style can be intricate, leading to the infamous “callback hell.” Node.js tries to simplify this with “promises,” a more functional approach.

A benefit of Node’s single-threaded nature is its simplicity. Even if callbacks from asynchronous calls like Service A and Service B can execute in any sequence, they run on the same thread, eliminating concurrency issues, like race condition, etc.

By sidestepping thread-based code flow, L is no longer bound by the OS’s scheduling limits but by hardware capacity and framework overhead. The need for additional hardware arises only when the current hardware is maxed out.

***

The functional style appears promising, but it diverges from the intuitive nature of the thread-per-request model. While the functional approach provides scalability and fault tolerance, its code is often less straightforward than that of the thread-per-request model. Yet, is it impossible to combine the best of both worlds: the scalability and fault tolerance of the functional style with the clarity of the thread-per-request model? Fortunately, it’s not. Enter Lightweight Threads.

Lightweight or User-mode Threads

So far, we have realised that a server capacity, denoted as L in Little’s formula, is often limited by its OS’s thread scheduling capability. Traditional threads, being a limited resource, had to be capped at some relatively small number, leading to the adoption of complex solutions like circuit breakers and functional programming style. But what if threads were more efficient?

Languages like Go and Erlang provide lightweight threads (Go’s goroutines, and Erlang’s processes). Similarly, the relatively new OpenJDK’s Project Loom introduces this functionality for Java 19 and later versions. These threads are managed by the language or library runtime, not the OS, allowing for more efficient scheduling.

Lightweight threads offer many advantages similar to traditional OS threads, including a straightforward control flow and the capability to block and wait for resources while permitting other threads to operate on the CPU. However, unlike traditional threads, lightweight threads are not managed by the OS, leading to quicker context-switching and reduced system resource consumption. Consequently, a single machine can efficiently manage millions of these threads. This offers a distinct advantage. The runtime, having a deeper understanding of the thread’s purpose and behaviour, can often outperform the OS in scheduling decisions. It recognizes that these threads operate in short bursts and frequently block, a behaviour not typical of heavyweight threads. This knowledge allows the runtime to employ an M:N scheduling model, where a larger number of lightweight threads (M) are mapped onto a smaller number of OS threads (N).

In essence, lightweight threads offer a path to maximize hardware efficiency without compromising code simplicity.

Conclusion

As the world continues its relentless march towards digital ubiquity, ensuring the scalability and fault tolerance of web applications becomes non-negotiable. While hardware advancements are laudable, the true bottleneck often lies in the software, especially the OS’s thread management capabilities. By embracing functional programming, lightweight threads, and a thorough understanding of Little’s Law, developers and system administrators can unlock the full potential of their servers, ensuring optimal performance even under duress.

References: this article is influenced by a series of presentations and articles by Ron Pressler (Tech lead, Project Loom)