How to measure the delay of receiving packets precisely or approximately
I’m trying to measure the latency when a packet enters the Rx buffer and is copied to application memory. I’m measuring it with this code :
struct timespec start, end;
clock_gettime(CLOCK_REALTIME, &start);
recvfrom(sock, msg, msg_len, 0, &client, &client_addrlen);
clock_gettime(CLOCK_REALTIME, &end);
I know this doesn’t measure latency precisely. However, I can calculate the average delay by receiving many packets, measuring each packet, and calculating them. Is there any way to measure latency more precisely? (For example, latency = (time to complete recvfrom() - (time from which the NIC received the packet)
).
For devices and device drivers, I’m using Mellanox connectx-3 and mlx4_en.
Solution
I was able to get almost exact numbers via recvmsg().
References
- The most helpful link ( https://stackoverflow.com/a/47329376/5215330 ).
- Network timestamp ( https://www.kernel.org/doc/Documentation/networking/timestamping.txt )
- Linux Precision Time Protocol ( http://linuxptp.sourceforge.net/ ).
Code
I’m reproducing the code in the first link. This code is not ready to run, but just a snippet of working code.
static struct timespec handle_time(struct msghdr *msg) {
struct cmsghdr *cmsg = CMSG_FIRSTHDR(msg);
struct scm_timestamping *ts = (struct scm_timestamping *)CMSG_DATA(cmsg);
return ts->ts[0];
}
...
char ctrl[64];
char *msg = malloc(64);
int val = SOF_TIMESTAMPING_RX_HARDWARE | SOF_TIMESTAMPING_RX_SOFTWARE
| SOF_TIMESTAMPING_SOFTWARE | SOF_TIMESTAMPING_RAW_HARDWARE;
setsockopt(sock_fd, SOL_SOCKET, SO_TIMESTAMPING, &val, sizeof(val));
user buffer
struct iovec iov = {
.iov_base = msg,
.iov_len = msg_len,
};
ancillary message header
struct msghdr m = {
.msg_name = &client_addr, // struct sockaddr_in
.msg_namelen = client_addrlen, // socklen_t
.msg_iov = &iov,
.msg_iovlen = 1,
.msg_control = &ctrl,
.msg_controllen = sizeof(ctrl),
};
while (1) {
memset(msg, 0, msg_len);
num_received = recvmsg(sock_fd, &m, 0);
start = handle_time(&m);
clock_gettime(CLOCK_REALTIME, &end);
if (verbose) {
double elapsed_time = time_diff(start, end) / 1000;
total_elapsed += elapsed_time;
count++;
printf("%f us %f us\n", elapsed_time, total_elapsed / count);
}
if (sendto(sock_fd, msg, msg_len, 0, (struct sockaddr *) &client_addr, client_addrlen) < 0) {
perror("\nMessage Send Failed\n");
fprintf(stderr, "Value of errno: %d\n", errno);
}
}
The key point is to use setsockopt()
and recvmsg().
The key mechanism is that when you set an option for a socket FD, the kernel sets a timestamp based on the timestamp flag. After you set them, if you receive a message with struct msghdr
, the kernel will audit the timestamp in SW or HW. When you view the data, you will be able to get 3 timestamps. This information can be interpreted as follows:
The structure can return up to three timestamps. This is a legacy
feature. At least one field is non-zero at any time. Most timestamps
are passed in ts[0]. Hardware timestamps are passed in ts[2]. ts[1] used to hold hardware timestamps converted to system time. Instead, expose the hardware clock device on the NIC directly as a HW PTP clock source, to allow time conversion in userspace and optionally synchronize system time with a userspace PTP stack such as linuxptp. For the PTP clock API, see Documentation/driver-api/ptp.rst.
For more information, see 2.1 in Documentation/networking/timestamping.txt
.
If you want to see the hardware timestamp, then you need to have a specific hardware (reference this comment) and change its characteristics with ioctl().
However, there is a handy tool called Linuxptp
that does the job.