Skip to content

Device

The Device class is the main entry point for interacting with the hardware device. It is used to allocate Ensō Pipes as well as to configure the device. It can also be used to efficiently receive data from multiple pipes, avoiding the need to probe each pipe individually.

Every I/O thread in a program should instantiate its own Device instance using the Device::Create() factory method. Device objects, like pipe objects, are not meant to be thread safe. They are designed to be used by a single thread only.

Allocating Ensō Pipes

After instantiating a device, the application can allocate Ensō Pipes of any of the three types, using the appropriate method:

RX Ensō Pipes and RX/TX Ensō Pipes can also be set as fallback when being allocated. Fallback pipes receive packets that do not match any explicit bind rules. See Binding and flow steering for more details on how the NIC steers packets to fallback pipes.

To set a pipe as fallback, both Device::AllocateRxPipe() and Device::AllocateRxTxPipe() accept an optional boolean argument that specifies whether the pipe should be set as fallback. For example:

RxPipe* rx_pipe_1 = dev->AllocateRxPipe();     // Normal pipe, non-fallback.
RxPipe* rx_pipe_2 = dev->AllocateRxPipe(true); // Set as fallback.

Receiving Data from Multiple Pipes

Threads can also use Device instances to figure out which pipe has data pending to be received. This is useful when the application needs to receive data from multiple pipes, as it avoids the need to probe each pipe individually.1

To figure out the next pipe with data pending, the application can call Device::NextRxPipeToRecv() or Device::NextRxTxPipeToRecv(). This will return the next pipe with data pending to be received, or nullptr if no pipe has data pending. Here is an example that receives data from multiple pipes:

// Allocate device.
std::unique_ptr<Device> dev = Device::Create();

// Allocate RX pipes.
RxPipe* rx_pipe_1 = dev->AllocateRxPipe();
RxPipe* rx_pipe_2 = dev->AllocateRxPipe();

while (keep_running) {
  // Figure out the next pipe with data pending to be received.
  RxPipe* pipe = dev->NextRxPipeToRecv();

  if (pipe == nullptr) continue;

  pipe->Recv(); // (1)!

  // Do something with the received data.
  // [...]

  pipe->Clear();
}
  1. ℹ Refer to the RX Ensō Pipe documentation for more information on how to receive data from an Ensō Pipe.

Note

There is an important caveat to consider when using these methods: they do not work if the application has a mix of RX and RX/TX pipes. If you plan to use those methods, make sure you only use one type of RX pipe.

Configuring the Device

You may also use a Device instance to configure the hardware device.

Hardware Rate Limiter

The hardware implementation includes a rate limiter that can be used to limit the rate at which packets are sent. The rate limiter is applied to all packets sent by the device, regardless of the pipe they are sent on. That means that even if you enable rate limiting from a specific thread, it will affect pipes from all threads. The rate limiter can be enabled by calling Device::EnableRateLimiting() and disabled by calling Device::DisableRateLimiting().

When enabling the rate limiter, you specify a fraction num / den of the maximum hardware flit rate (a flit is 64 bytes). The maximum hardware flit rate is defined by the kMaxHardwareFlitRate constant and is 200MHz by default. This will cause packets from all queues to be sent at a rate of num / den * kMaxHardwareFlitRate flits per second. Note that this is slightly different from how we typically define throughput and you will need to take the packet sizes into account to set this properly.

For example, suppose that you are sending 64-byte packets. Each packet occupies exactly one flit. For this packet size, line rate at 100Gbps is 148.8Mpps. So if kMaxHardwareFlitRate is 200MHz, line rate actually corresponds to a 744/1000 rate. Therefore, if you want to send at 50Gbps (50% of line rate), you can use a 372/1000 (or 93/250) rate.

The other thing to notice is that, while it might be tempting to use a large denominator in order to increase the rate precision. This has the side effect of increasing burstiness. The way the rate limiter is implemented, we send a burst of num consecutive flits every den cycles. Which means that if num is too large, it might overflow the receiver buffer. For instance, in the example above, 93/250 would be a better rate than 372/1000. And 3/8 would be even better with a slight loss in precision.

You can find the maximum packet rate for any packet size by using the expression: line_rate / ((pkt_size + 20) * 8). So for 100Gbps and 128-byte packets we have: 100e9 / ((128 + 20) * 8) packets per second. Given that each packet is two flits, for kMaxHardwareFlitRate = 200e6, the maximum rate is 100e9 / ((128 + 20) * 8) * 2 / 200e6, which is approximately 125/148. Therefore, if you want to send packets at 20Gbps (20% of line rate), you should use a 25/148 rate.

Hardware Time Stamping

The hardware implementation can also time stamp packets as they are sent and compute the RTT when they return. This is useful to measure latency. As with the rate limiter, this configuration is applied to all pipes. You can enable time stamping by calling Device::EnableTimeStamping() and disable it by calling Device::DisableTimeStamping().

When timestamping is enabled, all outgoing packets will receive a timestamp and all incoming packets will have an RTT (in number of cycles). You may use get_pkt_rtt() to retrieve the RTT for a returning packet. This function will return the RTT in number of cycles. You can convert it to nanoseconds by multiplying it by kNsPerTimestampCycle.

Round-Robin Steering

As described in Binding and flow steering, the NIC sends packets that do not match any binding rules to fallback pipes. By default, this is done using a hash of the five-tuple. You can change it to use round-robin instead by using Device::EnableRoundRobin() or revert back to the default by using Device::DisableRoundRobin().


  1. This is analogous to the select(2) system call in POSIX.