-
-
Notifications
You must be signed in to change notification settings - Fork 415
Performance Tuning
Contents
Assuming everything is going well we can focus on performance tuning.
Let's see the related possible settings at once first. These settings are used with 74/20 Mbps connection, 4 GB RAM and 1 local disk device and with rTorrent v0.9.6
.
# Global upload and download rate in KiB, `0` for unlimited (`download_rate`, `upload_rate`)
throttle.global_down.max_rate.set_kb = 8700
throttle.global_up.max_rate.set_kb = 2200
# Maximum number of simultaneous downloads and uploads slots (global slots!) (`max_downloads_global`, `max_uploads_global`)
throttle.max_downloads.global.set = 300
throttle.max_uploads.global.set = 300
# Maximum and minimum number of peers to connect to per torrent while downloading (`min_peers`, `max_peers`) Default: `100` and `200` respectively
throttle.min_peers.normal.set = 99
throttle.max_peers.normal.set = 100
# Same as above but for seeding completed torrents (seeds per torrent), `-1` for same as downloading (`min_peers_seed`, `max_peers_seed`) Default: `-1` for both
throttle.min_peers.seed.set = -1
throttle.max_peers.seed.set = -1
# Maximum number of simultaneous downloads and uploads slots per torrent (`max_uploads`) Default: `50` for both
throttle.max_downloads.set = 50
throttle.max_uploads.set = 50
# Set the numwant field sent to the tracker, which indicates how many peers we want.
# A negative value disables this feature. Default: `-1`
trackers.numwant.set = 100
# Set the max amount of memory address space used to mapping file chunks. This refers to memory mapping, not
# physical memory allocation. Default: `1GB` (`max_memory_usage`)
# This may also be set using ulimit -m where 3/4 will be allocated to file chunks.
pieces.memory.max.set = 2048M
# Maximum number of connections rtorrent can accept/make
network.max_open_sockets.set = 999
# Maximum number of open files rtorrent can keep open (you have to modify the system wide settings with ulimit!)
network.max_open_files.set = 600
# Maximum number of simultaneous HTTP request (used by announce or scrape requests) Default: `32`
network.http.max_open.set = 99
# Send and receive buffer size for socket. Disabled by default (`0`), this means the default is used by OS
# (you have to modify the system wide settings!)
# Increasing buffer sizes may help reduce disk seeking, connection polling as more data is buffered each time
# the socket is written to. It will result higher memory usage (not visible in rtorrent process!).
network.receive_buffer.size.set = 4M
network.send_buffer.size.set = 12M
# Preloading a piece of a file. Default: `0` Possible values: `0` (Off) , `1` (Madvise) , `2` (Direct paging).
pieces.preload.type.set = 2
#pieces.preload.min_size.set = 262144
#pieces.preload.min_rate.set = 5120
# TOS of peer connections. Default: `throughput`. If the option is set to `default` then the system default TOS
# is used. A hex value may be used for non-standard settings.
# Possible values: `[default|lowdelay|throughput|reliability|mincost]` or a hex value.
#network.tos.set = throughput
# CURL options to add skip TLS verification (for nonofficial SSL trackers)
network.http.ssl_verify_host.set = 0
network.http.ssl_verify_peer.set = 0
# CURL option to lower DNS timeout. Default: `60`.
network.http.dns_cache_timeout.set = 25
# Max packet size using xmlrpc. Default: `524288`
network.xmlrpc.size_limit.set = 2M
# Save all the sessions in every 12 hours instead of the default 20 minutes.
schedule2 = session_save, 1200, 43200, ((session.save))
# Prune file status in every 24 hours, this is the default setting.
#schedule2 = prune_file_status, 3600, 86400, ((system.file_status_cache.prune))
# Whether to allocate disk space for a new torrent. Default: `0`
#system.file.allocate.set = 0
rTorrent
uses a different philosophy than most other torrent clients.
slots
- can be upload or download slots
- determine how many peers
can actually transfer data at the same time, while rTorrent
can be connected with way more peers
. The allowed numbers of connected peers
should be 2 or 3 times higher than the allowed number of slots
.
-
throttle.max_downloads.global
,throttle.max_uploads.global
: maximum number of global simultaneous downloads and uploads slots. These values limit the global slots, can be seen at the right part of status bar:[U 45/300] [D 179/300]
-
throttle.max_uploads
,throttle.max_downloads
: maximum number of simultaneous downloads and uploads slots per torrent. It can be seen at the bottom left a torrent details page (using right arrow):Slots: U:0/50 D:0/50
-
throttle.max_peers.normal
,throttle.max_peers.seed
: maximum number of peers to connect to per torrent while downloading or seeding. It can be seen (along with the connected peers) at the bottom left a torrent details page (using right arrow):Peers: 43(0) Min/Max: 99/100
-
throttle.min_peers.normal
,throttle.min_peers.seed
: minimum number of peers to connect to per torrent while downloading or seeding. It can be seen (along with the connected peers) at the bottom left a torrent details page (using right arrow):Peers: 43(0) Min/Max: 99/100
The min_peers
values are responsible for asking more peers during an announce request. When the client has less than min_peers - (peer_list_size / 2)
connections for a download and PEX doesn't result in connection for a download (if it's enabled at all), it will attempt to request more from available trackers using tracker min interval
, otherwise using tracker interval
(note: this feature is broken in v0.9.6/0.13.6
). 30 seconds after a request the client will attempt another if more than 10 new peer connections were gained or less than 3 requests have been performed. Else it will try the next tracker group in the list, but not other trackers in the same group. This behavior should give enough peers while minimizing the number of tracker requests, although it will use somewhat longer time than other more aggressive clients. If these values are 0
then rTorrent uses tracker interval
all the time and won't ask for new peers from the given tracker, peers can still connect though (note: does this part works at all?).
It all depends on the global connection speeds (throttle.global_down.max_rate
, throttle.global_up.max_rate
) and available RAM.
-
Every upload slot should have got at least
5 KiB/s
speed left (it's not really problem anymore with fast connections nowadays). Taking the above example, in the worst case download slots have29 KiB/s
(8700/300
) and upload slots have7.3 KiB/s
(2200/300
). That means the number of the global download slot can be increased if we notice that we run out of it. -
throttle.max_downloads
andthrottle.max_uploads
slots are50
now. In the worst case this allows downloading and uploading6
(300/50
) torrents at the same time, but this usually not a problem for seeding. -
max_peers
settings for downloading and seeding should be at least 2 times higher than the number of slots per torrent, hence the value of100
for them. -
min_peers
settings are99
for both uploading and downloading, meaning we always want to ask the tracker for new peers more frequently, using trackermin interval
.
For every download or upload slot you allow you need a chunk's worth of RAM in the system to cache the chunks. E.g. if the current torrent your are uploading has a chunk size of 4 MiB
then you would need 200 MiB
(4 MiB * 50 slots
) of RAM for rTorrent
to use. Chunk sizes per torrent range from 256 KiB
to around 8 MiB
.
That means, we need 2.4 GiB
of RAM (4 MiB * 600 slots
) if the average chunk size is 4 MiB
in the worst case.
pieces.memory.max
is commonly misunderstood setting: it doesn't limit or set the amount of RAM rTorrent
can use (see Reduce disk usage section below) but limits the amount of memory address space used to mapping file chunks. This refers to memory mapping, not physical memory allocation. Default value is 1 GiB
That's how can happen that you never see e.g. greater value than 800 MiB
for RES
in htop
beside of the rtorrent
process.
For fast downloads and/or large number of peers this may quickly be exhausted causing the client to hang while it syncs to disk. You may increase this limit.
In theory, we can reduce disk i/o :) See this issue why: #443
The network.send_buffer.size
and network.receive_buffer.size
options can be used to adjust the socket send and receive buffer sizes. If you set these to a low number, you may see reduced throughput, especially for high latency connections. Increasing buffer sizes may help reduce disk seeking, connection polling as more data is buffered each time the socket is written to. See Networking tweaks section how to adjust it system wide.
It affects memory usage: this memory will not be visible in rtorrent
process, this sets the amount of kernel memory is used for your sockets. In low-memory, high-bandwidth case you still want decent buffer sizes. It is however the number of upload/downloading peers you want to reduce. (Reference: #435)
When a piece is to be uploaded to a peer it can preload the piece of the file before it does the non-blocking write to the network. This will not complete the whole piece if parts of the piece is not already in memory, having instead to try again later. (Reference: #418)
pieces.preload.type
Default: 0
. Possible values:
-
0
= off: it doesn't do any. -
1
= Madvise: it calls 'madvise' on the file for the specific mmap'ed memory range, which tells the kernel to load it in memory when it gets around to it. Which is hopefully before we write to the network socket. -
2
= Direct paging: we 'touch' each file page in order to force the kernel to load it into memory. This can help if you're dealing with very large number of peers and large/many files, especially in a low-memory setting, as you can avoid thrashing the disk where loaded file pages get thrown out before they manage to get sent.
Related settings:
#pieces.preload.min_size.set = 262144
#pieces.preload.min_rate.set = 5120
network.max_open_files
limits the maximum number of open files rTorrent
can keep open. By default rTorrent uses variable sized fd_set
's depending on the process sysconf(_SC_OPEN_MAX)
limit. See Networking tweaks section how to adjust it system wide.
Large fd_set
's cause a performance penalty as they must be cleared each time the client polls the sockets. When using select
or epoll
(until libcurl
is fixed) based polling use an open files limit that is reasonably low (is it still the case???). The widely used default of 1024
is enough for most users and 64
is minimum. Those with embedded devices or older platforms might need to set the limit much lower than the default.
(Due to libcurl
's use of fd_set
for polling, rTorrent
cannot at the moment move to a pure epoll
implementation. Currently the epoll
code uses select
based polling if, and only if, libcurl
is active. All non-libcurl
sockets are still in epoll
, but select
is used on the libcurl
and the epoll
-socket. (is it still the case???))
system.file.allocate
specifies whether to allocate disk space for a new torrent. Default is 0
.
Why would you care about this setting? Because "it’s beneficial in the as-less-fragmentation-as-possible sense".
In short: if you use btrfs, ext4, ocfs2, xfs file system you can enable this without having any performance impact.
When it's enabled (set to 1
):
- on a non-blocking file system that supports
fallocate
(like btrfs, ext4, ocfs2, xfs) that is always used to resize files, hence there is no performance issue (takes only a few microseconds to allocate space for huge files) - if
fallocate
isn't supported by a blocking file system and--with-posix-fallocate
was used during compilation oflibtorrent
then that is used to resize files, it can be significantly slower then the above one - on Mac OS X a different method is used
- if none of the above method is available then rtorrent falls back to disabled state, even when it's set to
1
- if priority of certain files of a download are set to
off
before starting it then file allocation isn't triggered for them
When it's disabled (set to 0
):
- Opening a torrent causes files to be created and resized with
ftruncate
(ftruncate
has problem on vfat filesystem, though, so another method is used in this case). This does not actually use disk space until data is written, despite what the file sizes are reported as. Usedu
without the flag--apparent-size
to see the real disk usage.
rTorrent
saves all the sessions in every 20 minutes by default. With large amount of torrents this can be a disk performance hog (see #180).
Increase this interval, e.g. to 12 hours instead, if you think you're system is stable and/or you don't care about the possible loss of resume and local stat in the meantime, with:
schedule2 = session_save, 1200, 43200, ((session.save))
When using using the load
commands, rtorrent caches the file's path and metadata, and avoids attempting to load any files that haven't changed. This is helpful when using watch directories that may have a large amount of torrent files that have already been loaded, but the only way to clear the cache is to clear it fully with the system.file_status_cache.prune
command. The default schedule of 24 hours is good enough for most cases, but you may wish to tweak the timing. Note that the command itself is very quick to run, but any subsequent load commands may be slow while rTorrent re-reads all the data from scratch.
schedule2 = prune_file_status, 3600, 86400, ((system.file_status_cache.prune))
Along with Name resolving enhancements we can disable the built-in DNS cache of curl
in favor of an external one with network.http.dns_cache_timeout.set=0
. Default is 60
sec.
To use higher settings for couple of the above settings the system wide limit should be raised for them.
network.max_open_files
is limited to ulimit -n
. If you want to increase it you can use ulimit -n new_value
or apply it permanently via /etc/security/limits.conf
on Ubuntu (default: 1024
):
#<domain> <type> <item> <value>
username soft nofile 10240
username hard nofile 10240
When tuning for large footprint installs, a large numbers for settings like slots, maximum file identifiers, and maximum sockets may cause issues (unexpected pausing and queuing of torrents) when using systemd. This is the case even when system wide limits (ulimit) have been changed to account for the added resource usage.
Systemd has implicit resource limits (4096) which don't obviously state their override of ulimit values. You will need to add LimitNOFILE=xxxx to either the rtorrent systemd definition or to /etc/systemd/system.conf to address this.
network.receive_buffer.size
, network.send_buffer.size
are limited to sysctl -a | grep -i rmem
and sysctl -a | grep -i wmem
. You have to change net.core.rmem_max
, net.ipv4.tcp_rmem
, net.core.wmem_max
, net.ipv4.tcp_wmem
in /etc/sysctl.conf
to the desired values along with some other tweaks:
# Maximum Socket Receive Buffer. 16MB per socket - which sounds like a lot, but will virtually never consume that much. Default: 212992
net.core.rmem_max = 16777216
# Maximum Socket Send Buffer. 16MB per socket - which sounds like a lot, but will virtually never consume that much. Default: 212992
net.core.wmem_max = 16777216
# Increase the write-buffer-space allocatable: min 4KB, def 12MB, max 16MB. Default: 4096 16384 4194304
net.ipv4.tcp_wmem = 4096 12582912 16777216
# Increase the read-buffer-space allocatable: min 4KB, def 12MB, max 16MB. Default: 4096 16384 4194304
net.ipv4.tcp_rmem = 4096 12582912 16777216
# Tells the system whether it should start at the default window size only for new TCP connections or also for existing TCP connections that have been idle for too long. Default: 1
net.ipv4.tcp_slow_start_after_idle = 0
# Allow reuse of sockets in TIME_WAIT state for new connections only when it is safe from the network stack’s perspective. Default: 0
net.ipv4.tcp_tw_reuse = 1
# Do not last the complete time_wait cycle. Default: 0
# only works from Linux 2.4 to 4.11
net.ipv4.tcp_tw_recycle = 1
# Minimum time a socket will stay in TIME_WAIT state (unusable after being used once). Default: 60
net.ipv4.tcp_fin_timeout = 30
rTorrent
sometimes can hang on hostname lookups, even with normal http/https requests. Here it is what we can do about it.
c-ares
is a C library for asynchronous DNS requests (including name resolves). Here you can find instructions how to compile it or just simply use rtorrent-ps by @pyroscope, its build script will do everything you need.
In addition to the above, it's strongly advised to use DNS caching either locally or for your whole network (e.g. on a OpenWRT powered router). Here you can find instructions how to set it up on Ubuntu.
- libtorrent manual (NB: This is not the rakshasa libtorrent, but still contains useful information)
- lost rTorrent docs
- rTorrent research
- c-ares
- Local DNS Cache
- Networking tweaks