2.1. Architectural Overview

On this page, we provide an overview of SimBricks’ architecture and how it connects existing simulators into an end-to-end virtual testbed. In the end, you should have all the required knowledge to extend SimBricks by adding a new Simulator or interface. Feel free to reach out to us if you have any questions or want to discuss an idea. Getting Help provides information on how to do so.

2.1.1. Connecting Simulators

We begin by introducing the high-level idea for connecting existing simulators. For the moment, assume we are working with a simple system made up of just two simulators, for example, a host simulator like gem5 that we want to connect to some simulator for a PCIe device.

In a real system, components connect through common interfaces like PCIe or Ethernet. Many simulators already expose an API for extending the simulation with components or devices that attach through these interfaces. SimBricks uses so-called adapters to emulate these, leveraging the provided API and forwarding events of interest to the adapter in the other connected simulator using the SimBricks protocol. This approach ensures modularity, or the ability to flexibly combine simulators into a virtual testbed. Adding adapters at common interfaces means that you can swap out one simulator for another without having to change the adapter at the other side of the connection under the assumption that the interface doesn’t change. For example, adding a PCIe adapter to our host simulator allows us to connect any simulator that implements a device-side PCIe adapter to it. This approach even works for closed-source simulators as long as they offer the necessary extensibility.

In the previous paragraph, we deliberately differentiated between host-side and device-side PCIe adapters. The reason is that, depending on the concrete interface, the two sides of a connection are not necessarily symmetric. In the case of PCIe, the host can initiate BAR and PCIe config reads or writes. The device, on the other hand, performs DMA and raises interrupts. This is an important consideration when implementing the adapter and defining the protocol used for the communication between them.

example of a modular SimBricks simulation experiment

Fig. 2.1 A modular SimBricks simulation experiment. We can connect different simulators into an end-to-end virtual testbed by adding SimBricks adapters at natural boundaries like PCIe and Ethernet. Adapters always communicate in pairs exchanging messages which contain events of interest. The SimBricks protocol is used for communication, which also offers special messages to synchronize the respective simulators if necessary.

2.1.2. SimBricks Protocol

So far, we only talked about connecting two simulators and omitted the details of the protocol the respective adapters are using to exchange events of interest. Even in more complex systems, simulators and therefore also their adapters, are always connected pair-wise. A single simulator can have multiple adapters though, which are connected to adapters in multiple other simulators. You can see this in action in Fig. 2.1.

Two adapters communicate over a pair of optimized shared memory queues, sending and polling SimBricks protocol messages containing information about events in a FIFO manner. From each adapter’s point of view, there is one send and receive queue, respectively. Each adapter polls for messages in its receive queue, which ensures the fastest simulation time as long as every simulator runs on its own physical CPU core.

For the messages itself, interface-specific protocols are defined on top of the base SimBricks protocol (lib/simbricks/base/proto.h#L118-L131). The base SimBricks protocol stores two important fields at fixed offsets within the header (first 64 bytes) of each message:

  • own_type - An integer identifying the type of message, required to correctly interpret the message when processing it later.

  • timestamp - Timestamp when the event occurred. Required when synchronizing to process the event at the correct time in the receiving simulator.

Aside from these fields, the header layout can be freely customized by the interface-specific protocol. Additionally, the message size is freely configurable per interface to accommodate payloads of arbitrary size.

Let’s walk through the memory protocol as a simple example. It is defined in the file lib/simbricks/mem/proto.h and is used for the simulation of memory disaggregated systems. The high-level idea of memory disaggregation is that the host’s memory resides somewhere externally, for example, it could be attached via the network. To simulate such a system, we introduced a simple memory simulator. In terms of the interface, the host issues read or write requests to the memory. For reads, the memory replies with the read value as the message’s payload. Writes can be either regular or posted. For the former, the memory replies with a completion message while posted writes are send-and-forget.

Notice that this interface is asymmetric, which means that we have to take into account the two directions when defining the different message types: host to memory (h2m) and memory to host (m2h). An interface doesn’t have to necessarily be asymmetric. SimBricks also comes with a protocol for Ethernet (lib/simbricks/network/proto.h). Here, both sides transmit Ethernet packets to the other side in a send-and-forget manner, which is symmetric.

To define the layout of the different message types for the memory or any other interface-specific protocol, we simply add a struct per type containing the fields. When an adapter receives a message, it selects the correct struct using the own_type field from the base protocol. For this, a unique integer to store in the own_type field is required per type. It mustn’t collide with those used in the base SimBricks protocol, collisions with other interface-specific protocols, however, are fine. Similarly, we must ensure to not overwrite the base protocol’s fields with something else.

2.1.3. Synchronization

So far, we only covered transmitting important events in one simulator to another. In case you are just interested in functional simulation, for example, for debugging or functional testing, you can run the whole virtual testbed unsynchronized for the lowest simulation time and skip this section. In order to get sensible end-to-end measurements, however, you need to synchronize the advancement of virtual time between the simulators.

Synchronization in the case of SimBricks means informing the connected simulator that there will be no more messages to process up to some concrete timestamp. For this, the base protocol defines a special synchronization message type. Synchronization messages are sent over the same pair of send and receive queues as the interface-specific messages. However, sending these for every tick of a simulator’s virtual clock doesn’t scale. We can use some of SimBricks’ properties to reduce their number. First, we don’t need to synchronize simulators globally. Instead, it suffices to only do so pair-wise along the connections between adapters. In particular, this means that we don’t have to synchronize simulators that aren’t directly connected.

Furthermore, all messages are inserted into the shared memory queues in FIFO order of when their respective event occurred in the sending simulator. This guarantees that when polling the messages on the receiver side, the timestamps always increase monotonically. We use this together with the observation that links between components in real systems always have some latency to provide synchronization slack. Essentially, if one side of the connection polls a message with time t, it can safely advance to timestamp t + link latency. The link latency is configured by the user.

The link latency also helps with the frequency of synchronization messages. If we already sent a synchronization message containing t, then it suffices to only send another one when our local clock reaches t + link latency since the connected simulator, due to the link latency, couldn’t process any message from us in the meantime anyway. For accurate simulation, it therefore suffices to periodically send synchronization messages with the link latency as the period.

There is one last optimization. Every message carries a timestamp and can therefore serve as an implicit synchronization message. Whenever we send a message at time t, we can therefore reschedule sending a synchronization message to t + link latency. Depending on the expected frequency of messages, rescheduling may be more expensive than just sending the synchronization message periodically. This is, for example, the case for gem5.