udp-sender

Langue: en

Version: March 19, 2006 (debian - 07/07/09)

Section: 1 (Commandes utilisateur)

NAME

udp-sender - broadcast file on a LAN

SYNOPSIS

udp-sender [--file file] [--full-duplex] [--half-duplex] [--pipe pipe] [--portbase portbase] [--blocksize size] [--interface net-interface] [--mcast-addr data-mcast-address] [--mcast-all-addr mcast-all-address] [--max-bitrate bitrate] [--pointopoint] [--async] [--log file] [--min-slice-size min] [--max-slice-size max] [--slice-size] [--ttl time-to-live] [--fec stripesxredundancy/stripesize] [--print-seed] [--rexmit-hello-interval interval] [--autostart autostart] [--broadcast] [--min-clients clients] [--min-wait sec] [--max-wait sec] [--nokbd] [--retriesUntilDrop n]

DESCRIPTION

"Udp-sender" is used to broadcast a file (for instance a disk image) to multiple "udp-receivers" on the local LAN. In order to do this, it uses Ethernet multicast or broadcast, so that all receivers profit from the same physical datastream. Thus, sending to 10 destinations does not take more time than it would take to send just 2.

OPTIONS


Basic options

--file file
Reads data to be transmitted from file. If this parameter is not supplied, data to be transmitted is read from stdin instead.
--pipe command
Sends data through pipe before transmitting it. This is useful for compressing/decompressing it, or for stripping out unused blocks. The command gets a direct handle on the input file or device, and thus may seek inside it, if needed. "Udpcast" itself also keeps a handle on the file, which is used for an informal progress display. The command's stdout is a pipe to udpcast.
--log file
Logs some stuff into file.
--autostart n
Starts transmission after n retransmissions of hello packet, without waiting for a key stroke. Useful for unattended operation, where udp-sender is started from a cron-job for a broadcast/multicast at a scheduled time.

Networking options

The following networking options should be supplied both on the sender and the receivers:

--portbase portbase
Default ports to use for udpcast. Two ports are used: portbase and portbase+1 . Thus, Portbase must be even. Default is 9000. The same portbase must be specified for both "udp-sender" and "udp-receiver".
--interface interface
Network interface used to send out the data. Default is "eth0"
--broadcast
Use Ethernet broadcast, rather than multicast. Useful if you have Ethernet cards which don't support multicast.

By default, "udpcast" uses multicast. This allows sending the data only to those receivers that requested it. Ethernet cards of machines which don't participate in the transmission automatically block out the packets at the hardware level. Moreover, network switches are able to selectively transmit the packets only to those network ports to which receivers are connected. Both features thus allow a much more efficient operation than broadcast. This option should only be supplied on the sender.

--ttl time to live
Sets the time-to-live parameter for the multicast packets. Should theoretically allow to use UDPCast beyond the local network, but not tested for lack of a multicast router.
--mcast-all-addr address
Uses a non-standard multicast address for the control connection (which is used by the sender and receivers to ``find'' each other). This is not the address that is used to transfer the data.

By default "mcast-all-addr" is the Ethernet broadcast address if "ttl" is 1, and 224.0.0.1 otherwise. This setting should not be used except in very special situations, such as when 224.0.0.1 cannot be used for policy reasons.

The following networking options should be supplied only on the sender:

--mcast-addr address
Uses the given address for multicasting the data. If not specified, the program will automatically derive a multicast address from its own IP (by keeping the last 27 bits of the IP and then prepending 232).
--pointopoint
Point-to-point mode. Only a single receiver is allowed, but the data will be directly send to this receiver (in unicast mode), rather than multicast/broadcast all over the place. If no async mode is chosen, and there happens to be only one receiver, point-to-point is activated automatically.
--nopointopoint
Don't use point-to-point, even if there is only one single receiver.
--full-duplex
Use this option if you use a full-duplex network. T-base-10 or 100 is full duplex if equipped with a switch. Hub based networks, or T-base-2 networks (coaxial cable) are only half-duplex and you should not use this option with these networks, or else you may experience a 10% performance hit.

N.B. On high-latency WAN links, the full-duplex option can lead to substantial performance improvements, because it allows udp-sender to send more data while it is still waiting for the previous batch to get acknowledged.

--half-duplex
Use half duplex mode (needed for Hub based or T-base-2 networks). This is the default behavior in this version of udpcast.
-b blocksize
Choses the packet size. Default (and also maximum) is 1456.

Unidirectional mode (without return channel)

The options described below are useful in situations where no ``return channel'' is available, or where such a channel is impractical due to high latency. In an unidirectional setup (i.e. without return channel), the sender only sends data but doesn't expect any reply from the receiver.

Unidirectional options must be used together, or else the transfer will not work correctly. You may for example use the following command line:

"udp-sender --async --max-bitrate 10m --fec 8x8"

--async
Asynchronous mode. Do not request confirmations from the receiver. Best used together with forward error correction and bandwidth limitation, or else the receiver will abort the reception as soon as it misses a packet. When the receiver aborts the reception in such a way, it will print a list of packets lost in the slice causing the problem. You can use this list to tune the forward error correction parameters.
--max-bitrate bitrate
Limits bandwidth used by udpcast. Useful in asynchronous mode, or else the sender may send too fast for the switch and/or receiver to keep up. Bitrate may be expressed in bits per second (--bitrate 5000000), kilobits per second ("--bitrate 5000k") or megabits per second ("--bitrate 5m"). This is the raw bitrate, including packet headers, forward error correction, retransmissions, etc. Actual payload bitrate will be lower.
--fec interleave"x"redundancy"/"stripesize
Enables forward error correction. The goal of forward error correction is to transmit redundant data, in order to make up for packets lost in transit. Indeed, in unidirectional mode, the receivers have no way of requesting retransmission of lost packets, thus the only way to address packet loss is to include redundant information to begin with. The algorithm is designed in such a way that if r redundant packets are transmitted, that those can be used to compensate for the loss of any r packets in the same FEC group (stripe).

In order to increase robustness of the FEC algorithm against burst packet losses, each slice is divided in interleave stripes. Each stripe has stripesize blocks (if not specified, stripesize is calculated by diving slice-size by interleave). For each stripe, redundancy FEC packets are added. Stripes are organized in such a fashion that consecutive packets belong to different stripes. This way, we ensure that burst losses affect different stripes, rather than using all FEC packets of a single stripe. Example: "--fec 8x8/128"

--rexmit-hello-interval timeout
If set, rebroadcasts the HELLO packet used for initiating the casting each timeout milliseconds.

This option is useful together with asyc mode, because with async mode the receiver won't send a connection request to the sender (and hence won't get a connection reply). In async mode, the receivers get all needed information from the hello packet instead, and are thus particularly dependant on the reception of this packet, makeing retransmission useful.

This option is also useful on networks where packet loss is so high that even with connection requests, sender and receiver would not find each other otherwise.

--retriesUntilDrop retries
How many time to send a REQACK until dropping a receiver. Lower retrycounts make "udp-sender" faster to react to crashed receivers, but they also increase the probability of false alerts (dropping receivers that are not actually crashed, but merely slow to respond for whatever reason)

Keyboardless mode

The options below help to run a sender in unattended mode.

--min-clients n
Automatically start as soon as a minimal number of clients have connected.
--min-wait t
Even when the necessary amount of clients do have connected, still wait until t seconds since first client connection have passed.
--max-wait t
When not enough clients have connected (but at least one), start anyways when t seconds since first client connection have pased.
--nokbd
Do not read start signal from keyboard, and do not display any message telling the user to press any key to start.

Example:

"udp-sender -f zozo --min-clients 5 --min-wait 20 --max-wait 80"

If one client connects at 18h00.00, and 4 more within the next 5 minutes, start at 18h00.20. (5 clients connected, but min-wait not yet pased)
If one client connects at 18h00.00, and 3 more within the next 5 minutes, then a last one at 18h00.25, start right after.
If one client connects at 18h00.00, then 3 more within the next 15 minutes, then no one, start at 18h01.20. (not enough clients, but we start anyways after max-wait).

Tuning options (sender)

The following tuning options are all about slice size. Udpcast groups its data in slices, which are a series of blocks (UDP packets). These groups are relevant for

data retransmission: after each slice, the server asks the receivers whether they have received all blocks, and if needed retransmits what has been missing
forward error correction: each slice has its set of data blocks, and matching FEC blocks.
--min-slice-size size
minimum slice size (expressed in blocks). Default is 16. When dynamically adjusting slice size (only in non-duplex mode), never use smaller slices than this. Ignored in duplex mode (default).
--max-slice-size size
maximum slice size (expressed in blocks). Default is 1024. When dynamically adjusting slice size (only in non-duplex mode), never use larger slices than this. Ignored in duplex mode (default).
--default-slice-size size
Slice size used (starting slice size in half-duplex mode).

Tuning the forward error correction

There are three parameters on which to act:

redundancy
This influences how much extra packets are included per stripe. The higher this is, the more redundancy there is, which means that the transmission becomes more robust against loss. However, CPU time necessary is also proportional to redundancy (a factor to consider on slow PC's), and of course, a higher redundancy augments the amount of data to be transmitted.
interleave
This influences among how many stripes the data is divided. Higher interleave improves robustness against burst loss (for example, 64 packets in a row...). It doesn't increase robustness against randomly spread packet loss. Note: interleave bigger than 8 will force a smaller stripesize, due to the fact that slicesize is limited to 1024.
stripesize
How many data blocks there are in a stripe. Due to the algorithm used, this cannot be more than 128. Reducing stripe size is an indirect way of augmenting (relative) redundancy, without incurring the CPU penalty of larger (absolute) redundancy. However, a larger absolute redundancy is still preferable over a smaller stripesize, because it improves robustness against clustered losses. For instance, if 8/128 is preferable over 4/64, because with 8/128 the 8 FEC packets can be used to compensate for the loss of any of the 128 data packets, whereas with 4/64, each group of 4 FEC packets can only be used against its own set of 64 data packets. If for instance the first 8 packets were lost, they would be recoverable with 8/128, but not with 4/64.

Considering these, change parameters as follows:

If you observe long stretches of lost packets, increase interleave
If you observe that transfer is slowed down by CPU saturation, decrease redundancy and stripesize proportionnally.
If you observe big variations in packet loss rate, increase redundancy and stripesize proportionnally.
If you just observe high loss, but not necessarily clustered in any special way, increase redundancy or decrease stripesize
Be aware that network equipment or the receiver may be dropping packets because of a bandwidth which is too high. Try limiting it using "max-bitrate"
The receiver may also be dropping packets because it cannot write the data to disk fast enough. Use hdparm to optimize disk access on the receiver. Try playing with the settings in "/proc/sys/net/core/rmem_default" and "/proc/sys/net/core/rmem_max", i.e. setting them to a higher value.

SEE ALSO

udp-receiver

AUTHOR

Alain Knaff