Performance Tuning

Most users will probably not have a need to tune and optimse their Unbound installation, but it could be useful for large resolver installations. This page contains recommendations based on user feedback. If you have different experiences or have recommendations, please share them on the Unbound users mailing list.

Configuration

Set num-threads equal to the number of CPU cores on the system. For example, for 4 CPUs with 2 cores each, use 8.

On Linux, set so-reuseport: yes, that will significantly improve UDP performance (on kernels that support it, otherwise it is inactive, the unbound-control status command shows if it is active).

Set *-slabs to a power of 2 close to the num-threads value. Do this for msg-cache-slabs, rrset-cache-slabs, infra-cache-slabs and key-cache-slabs. This reduces lock contention.

Increase the memory size of the cache. Use roughly twice as much rrset cache memory as you use msg cache memory. For example, rrset-cache-size: 100m and msg-cache-size: 50m. Due to malloc overhead, the total memory usage is likely to rise to double (or 2.5x) the total cache memory that is entered into the config.

Set the outgoing-range to as large a value as possible, see the sections below on how to overcome the limit of 1024 in total. This services more clients at a time. With 1 core, try 950. With 2 cores, try 450. With 4 cores try 200. The num-queries-per-thread is best set at half the number of the outgoing-range, but you would like a whole lot to be able to soak up a spike in queries. Because of the limit on outgoing-range thus also limits num-queries-per-thread, it is better to compile with libevent (see the section below), so that there is no more 1024 limit on outgoing-range.

Set so-rcvbuf to a larger value (4m or 8m) for a busy server. This sets the kernel buffer larger so that no messages are lost in spikes in the traffic. Adds extra 9s to the reply-reliability percentage. The OS caps it at a maximum, on Linux, Unbound needs root permission to bypass the limit, or the admin can use sysctl net.core.rmem_max. On BSD change kern.ipc.maxsockbuf in /etc/sysctl.conf.

On OpenBSD change header and recompile kernel. On Solaris ndd -set /dev/udp udp_max_buf 8388608.

Also set so-sndbuf to a larger value (4m or 8m) for a busy server. Same as rcvbuf, but now for spikes in replies, and it is net.core.wmem_max. Might need a smaller value, as spikes are less common in replies, you can see rcv and snd buffer overruns with netstat -su, RcvbufErrors and SndbufErrors, and similar reports on BSD.

For the TCP listen backlog on Linux, it is possible to tweak the kernel parameters to allow larger values. Unbound attempts to increase this to enable it to handle spikes in incoming TCP or TLS connections. The number that unbound attempts is defined in TCP_BACKLOG in services/listen_dnsport.c, it does not need to be changed if the current value, about 256, is sufficient for you. However, the Linux kernel limits this value silently to a maximum configured into the kernel settings. The kernel can be tweaked to enable a higher number with net.core.somaxconn = 256 and net.ipv4.tcp_max_syn_backlog = 256.

Here is a short summary of optimisation config:

# some optimisation options.
server:
    # use all CPUs
    num-threads: <number of cores>

    # power of 2 close to num-threads
    msg-cache-slabs: <same>
    rrset-cache-slabs: <same>
    infra-cache-slabs: <same>
    key-cache-slabs: <same>

    # more cache memory, rrset=msg*2
    rrset-cache-size: 100m
    msg-cache-size: 50m

    # more outgoing connections
    # depends on number of cores: 1024/cores - 50
    outgoing-range: 950

    # Larger socket buffer.  OS may need config.
    so-rcvbuf: 4m
    so-sndbuf: 4m

    # Faster UDP with multithreading (only on Linux).
    so-reuseport: yes

The default setup works fine, but when a large number of users have to be served, the limits of the system are reached. Most pressing is the number of file descriptors, the default has a limit of 1024. To use more than 1024 file descriptors, use libevent or the forked operation method. These are described in sections below.

Using Libevent

Libevent is a BSD licensed cross platform wrapper around platform specific event notification system calls. Unbound can use it to efficiently use more than 1024 file descriptors. Install libevent (and libevent-devel, if it exists) with your favorite package manager. Before compiling unbound run ./configure --with-libevent.

Now you can give any number you like for outgoing-range. Also increase the num-queries-per-thread value.

# with libevent
outgoing-range: 8192
num-queries-per-thread: 4096

Users report that libevent-1.4.8-stable works well. Users have confirmed it works well on Linux and FreeBSD with 4096 or 8192 as values. Double the num-queries-per-thread and use that as outgoing-range.

Stable(old) distributions can package older versions (such as libevent-1.1), for which there are crashreports, thus you may need to upgrade your libevent. In unbound 1.2.0 a race condition in the libevent calls was fixed.

Unbound can compile from the libevent or libev build directory to make this easy; configure --with-libevent=/home/user/libevent-1.4.8-stable, or configure --with-libevent=/home/user/libev-3.52.

Note

If you experience crashes anyway, then you can try the following. Update libevent. If the problem persists, libevent can be made to use different system-call back-ends by setting environment variables. Unbound reports the back-end in use when verbosity is at level 4. By setting EVENT_NOKQUEUE, EVENT_NODEVPOLL, EVENT_NOPOLL, EVENT_NOSELECT, EVENT_NOEPOLL or EVENT_NOEVPORT to yes in the shell before you start unbound, some back-ends can be excluded from use. The poll(2) backend is reliable, but slow.

Forked Operation

Unbound has a unique mode where it can operate without threading. This can be useful if libevent fails on the platform, for extra performance, or for creating walls between the cores so that one cannot poison another.

To compile for forked operation, before compilation use ./configure --without-pthreads --without-solaris-threads to disable threads and enable forked operation. Because no locking has to be done, the code speeds up (about 10 to 20%).

In the config file, num-threads still specifies the number of cores you want to use (even though it uses processes and not threads). And note that the outgoing-range and cache memory values are all per thread. This means that much more memory is used, as every core uses its own cache. Because every core has its own cache, if one gets cache poisoned, the others are not affected.

# with forked operation
server:
    # use all CPUs
    num-threads: <number of cores>

    msg-cache-slabs: 1
    rrset-cache-slabs: 1
    infra-cache-slabs: 1
    key-cache-slabs: 1

    # more cache memory, rrset=msg*2
    # total usage is 150m*cores
    rrset-cache-size: 100m
    msg-cache-size: 50m

    # does not depend on number of cores
    outgoing-range: 950
    num-queries-per-thread: 512

    # Larger socket buffer.  OS may need config.
    so-rcvbuf: 4m

Because every process is using at most 1024 file descriptors now, the effective maximum is the number of cores * 1024. The config above uses 950 per process, for 4 processes gives a respectable 3800 sockets. The number of queries per thread is half the number of sockets, to guarantee that every query can get a socket, and some to spare for queries-for-nameservers.

Using forked operation together with libevent is also possible. It may be useful to force the OS to service the filedescriptors for different processes, instead of threads. This may have (radically) different performance if the underlying network stack uses (slow) lookup structures per-process.