54. Writing Efficient Code

54.1.1. Memory Copy: Do not Use libc in the Data Plane

Many libc functions are available in the DPDK, via the Linux* application environment. This can ease the porting of applications and the development of the configuration plane. However, many of these functions are not designed for performance. Functions such as memcpy() or strcpy() should not be used in the data plane. To copy small structures, the preference is for a simpler technique that can be optimized by the compiler. Refer to the VTune™ Performance Analyzer Essentials publication from Intel Press for recommendations.

For specific functions that are called often, it is also a good idea to provide a self-made optimized function, which should be declared as static inline.

The DPDK API provides an optimized rte_memcpy() function.

54.1.2. Memory Allocation

Other functions of libc, such as malloc(), provide a flexible way to allocate and free memory. In some cases, using dynamic allocation is necessary, but it is really not advised to use malloc-like functions in the data plane because managing a fragmented heap can be costly and the allocator may not be optimized for parallel allocation.

If you really need dynamic allocation in the data plane, it is better to use a memory pool of fixed-size objects. This API is provided by librte_mempool. This data structure provides several services that increase performance, such as memory alignment of objects, lockless access to objects, NUMA awareness, bulk get/put and per-lcore cache. The rte_malloc () function uses a similar concept to mempools.

54.1.3. Concurrent Access to the Same Memory Area

Read-Write (RW) access operations by several lcores to the same memory area can generate a lot of data cache misses, which are very costly. It is often possible to use per-lcore variables, for example, in the case of statistics. There are at least two solutions for this:

• Use RTE_PER_LCORE variables. Note that in this case, data on lcore X is not available to lcore Y.
• Use a table of structures (one per lcore). In this case, each structure must be cache-aligned.

Read-mostly variables can be shared among lcores without performance losses if there are no RW variables in the same cache line.

54.1.4. NUMA

On a NUMA system, it is preferable to access local memory since remote memory access is slower. In the DPDK, the memzone, ring, rte_malloc and mempool APIs provide a way to create a pool on a specific socket.

Sometimes, it can be a good idea to duplicate data to optimize speed. For read-mostly variables that are often accessed, it should not be a problem to keep them in one socket only, since data will be present in cache.

54.1.5. Distribution Across Memory Channels

Modern memory controllers have several memory channels that can load or store data in parallel. Depending on the memory controller and its configuration, the number of channels and the way the memory is distributed across the channels varies. Each channel has a bandwidth limit, meaning that if all memory access operations are done on the first channel only, there is a potential bottleneck.

By default, the Mempool Library spreads the addresses of objects among memory channels.

54.1.6. Locking memory pages

The underlying operating system is allowed to load/unload memory pages at its own discretion. These page loads could impact the performance, as the process is on hold when the kernel fetches them.

To avoid these you could pre-load, and lock them into memory with the mlockall() call.

if (mlockall(MCL_CURRENT | MCL_FUTURE)) {
    RTE_LOG(NOTICE, USER1, "mlockall() failed with error \"%s\"\n",
            strerror(errno));
}

54.3.1. Lower Packet Latency

Traditionally, there is a trade-off between throughput and latency. An application can be tuned to achieve a high throughput, but the end-to-end latency of an average packet will typically increase as a result. Similarly, the application can be tuned to have, on average, a low end-to-end latency, at the cost of lower throughput.

In order to achieve higher throughput, the DPDK attempts to aggregate the cost of processing each packet individually by processing packets in bursts.

Using the testpmd application as an example, the burst size can be set on the command line to a value of 16 (also the default value). This allows the application to request 16 packets at a time from the PMD. The testpmd application then immediately attempts to transmit all the packets that were received, in this case, all 16 packets.

The packets are not transmitted until the tail pointer is updated on the corresponding TX queue of the network port. This behavior is desirable when tuning for high throughput because the cost of tail pointer updates to both the RX and TX queues can be spread across 16 packets, effectively hiding the relatively slow MMIO cost of writing to the PCIe* device. However, this is not very desirable when tuning for low latency because the first packet that was received must also wait for another 15 packets to be received. It cannot be transmitted until the other 15 packets have also been processed because the NIC will not know to transmit the packets until the TX tail pointer has been updated, which is not done until all 16 packets have been processed for transmission.

To consistently achieve low latency, even under heavy system load, the application developer should avoid processing packets in bunches. The testpmd application can be configured from the command line to use a burst value of 1. This will allow a single packet to be processed at a time, providing lower latency, but with the added cost of lower throughput.

54.5.1. Inline Functions

Small functions can be declared as static inline in the header file. This avoids the cost of a call instruction (and the associated context saving). However, this technique is not always efficient; it depends on many factors including the compiler.

54.5.2. Branch Prediction

The Intel® C/C++ Compiler (icc)/gcc built-in helper functions likely() and unlikely() allow the developer to indicate if a code branch is likely to be taken or not. For instance:

if (likely(x > 1))
    do_stuff();