Whenever I download something it starts slow but gets faster overtime then stays the same whereas speed in UDP increases or decreases randomly.
So my question is, what causes TCP to get faster and why UDP is more unstable compared to TCP?
So my question is what causes TCP to get faster
Most likely that is due to TCP's slow start feature, which is designed to avoid overloading the network with too much traffic by starting out at a conservative pace, and only increasing the transfer rate once the algorithm has recognized that the network is handling the initial rate without too many dropped packets. The rate will increase until packets start being dropped, at which point the TCP layer will back off a bit, until finally it (hopefully) arrives at the fastest transfer rate that network conditions can reliably support.
why UDP is too unstable when compared to TCP?
Unlike TCP, UDP doesn't make any attempt at congestion control; that sort of thing is left entirely up to the application programmer. All UDP does is send individual UDP packets (when the calling program asks it to by calling send() or sendto()), which may or may not arrive at their destination; if they do not arrive (for whatever reason) no further action is taken by the transport layer. So any particular behavior you see with UDP packets is more an indication of how your UDP-using application was programmed, than an indication of how the UDP transport layer behaves.
I have read some articles online and I got a pretty good idea about the TCP and UDP in general. However, I still have some doubts which I am sure not completely clear to me.
What is the proper way to calculate throughput ?
(Can't we just divide Total number of bytes received by total time taken ?)
What is that key feature in TCP that makes it have much much higher
throughput than UDP ?
UPDATE:
I understood that TCP uses windows which is nothing but that much segments can be sent before actually waiting for Acknowledgements. But my doubt is that in UDP segments are continuously sent without even bothering about Acknowledgements. So there is no extra overheads in UDP. Then, why the throughput of TCP is much much higher than that of UDP ?
Lastly,
Is this true ?
TCP throughput = (TCP Window Size / RTT) = BDP / RTT = (Link Speed in Bytes/sec * RTT)/RTT = Link Speed in Bytes/sec
If so then TCP throughput is always equals to the Know Link speed. And since the RTTs cancels out each other, the TCP throughput does not even depends on RTT.
I have seen in some network analysis tools like iperf, passmark performance test etc. that the TCP/UDP Throughput changes with Block size.
How is throughput dependent on Block size ?
Is Block size equals TCP window or UDP datagram size ?
What is the proper way to calculate throughput?
There are multiple ways, depending on what exactly you want to measure. They all boil down to dividing some number of bits (or bytes) to some duration, as you mention; what varies is which bits you are counting or (more rarely) which moments of time you are considering for measuring the duration.
The factors you need to take into account are:
At which layer in the network stack are you measuring throughput?
If you measure at the application layer, all that matters is what useful data you transmit to the other endpoint. For example, if you are transferring a file of 6 kB, the amount of data you count when measuring throughput is 6 kB (that is 6,000 bytes, not bits, and note the multiplier of 1000, not 1024; these conventions are common in networking).
This is usually called goodput and it may be different from what is actually sent at the transport layer (as in TCP or UDP), for two reasons:
1. Overhead due to headers
Each layer in the network adds a header to the data that introduces some overhead due to its transmission time. Moreover, the transport layer breaks your data into segments; this is because the network layer (as in IPv4 or IPv6) has a maximum packet size called MTU, typically 1,500 B in Ethernet networks. This value includes the network layer header size (e.g. the IPv4 header, which is variable in length but usually 20 B long) and the transport layer header (for TCP, it is also variable in length but usually 40 B long). This leads to a maximum segment size MSS (number of data bytes, without headers, in one segment) of 1500 - 40 - 20 = 1440 bytes.
Thus if we want to send 6 kB of application-layer data, we must break it into 6 segments, 4 of 1440 bytes each and one of 240 bytes. However at the network layer we end up sending 6 packets, 4 of 1500 bytes each and one of 300 bytes, for a total of 6.3 kB.
Here I have not considered the fact that the link layer (as in Ethernet) adds its own header and possibly also a suffix, which increases the overhead further. For Ethernet this is 14 bytes for the Ethernet header, optionally 4 bytes for VLAN tag, then a CRC of 4 bytes and a gap of 12 bytes, for a total of 36 bytes per packet.
If you consider a fixed-rate link, say of 10 Mb/s, depending on what you measure you will get a different throughput. Normally you want one of these:
The goodput, i.e. application layer throughput, if what you want to measure is application performance. For this example, you divide 6 kB by the transfer duration.
The link-layer throughput, if what you want to measure is network performance. For this example, you divide 6 kB + TCP overhead + IP overhead + Ethernet overhead = 6.3 kB + 5 * 36 B = 6516 B by the transfer duration.
Retransmission overheads
The Internet is a best-effort network, meaning that the packets will be delivered if possible, but may also be dropped. Packet drops are corrected by the transport layer, in case of TCP; for UDP, there is no such mechanism, which means that either the application does not care if some parts of the data do not get delivered, or the application implements retransmission itself on top of UDP.
Retransmission reduce goodput for two reasons:
a. Some data needs to be sent again, which takes time. This introduces a delay which is inversely proportional to the rate of the slowest link in the network between the sender and the receiver (a.k.a the bottleneck link).
b. Detecting that some data was not delivered needs feedback from the receiver to the sender. Due to propagation delays (sometimes called latency; caused by the finite speed of light in the cable), feedback can only be received by the sender with some latency, which slows down the transmission even more. In most practical cases, this is the most significant contribution to the extra delay caused by the retransmission.
Clearly, if you use UDP instead of TCP and you do not care about packet loss, you will of course get better performance. But for many applications, data loss cannot be tolerated, so such a measurement is meaningless.
There are some applications that do use UDP for transferring data. One is BitTorrent, which may use either TCP or a protocol they designed called uTP, which emulates TCP on top of UDP, but aims at being more efficient with many parallel connections. Another transport protocol implemented over UDP is QUIC, which also emulates TCP and offers multiplexing multiple parallel transfers over a single connection, and forward error correction to reduce retransmissions.
I will discuss forward error correction a little since it is related to your question about throughput. A naive way of implementing it is by sending every packet twice; in case one gets lost, the other still has a chance of being received. This reduces the amount of retransmissions to half, but also halves your goodput since you send redundant data (note that the network or link layer throughput remains the same!). In some cases this is fine; especially if the latency is very large, such as on intercontinental or satellite links. Moreover, some mathematical methods exist where you don't have to send a full copy of the data; for instance for every n packets you send, you send another reduntant one which is the XOR (or some other arithmetic operation) of them; if the redundant one gets lost, it doesn't matter; if one of the n packets gets lost, you can reconstruct it based on the redundant one and the other n-1. You can thus configure the overhead introduced by forward error correction to whatever amount of bandwidth you can spare.
How you are measuring the transfer time
Is the transfer completed when the sender finished sending the last bit over the wire, or does it also include the time it takes for the last bit to travel to the receiver? Additionally, does it include the time it takes to get a confirmation from the receiver, stating that all data has been received successfully and no retransmission is neede?
It really depends on what you want to measure. Note that for large transfers, one extra round-trip-time is insignificant in most cases (unless you are communicating, for instance, with a probe on Mars).
What is that key feature in TCP that makes it have much much higher throughput than UDP?
This is not true, although a common misconception.
In addition to retransmitting data when needed, TCP will also adjust its sending rate so that it will not cause packet drops by congesting the network. The adjustment algorithm has been perfected over decades, and usually converges quickly to the maximum rate supported by the network (actually, the bottleneck link). For this reason it is usually difficult to beat TCP in throughput.
With UDP, there is no rate limiting at the sender. UDP lets the application send as much as it wants. But if you try to send more than the network can handle, some of the data will be dropped, lowering your throughput, and also making the admin of the network you are congesting very angry. This means that sending UDP traffic at high rates is impractical (unless the goal is to DoS a network).
Some media applications are using UDP but rate-limiting the transfer at the sender at a very small rate. This is typically used in VoIP applications or Internet Radio, where you require very little throughput but low latency. I suppose this is one of the reasons for the misconception that UDP is slower than TCP; that is not the case, UDP can be as fast as the network allows.
As I said before, there are protocols such as uTP or QUIC, implemented over UDP, which achieve performance similar to TCP.
Is this true ?
TCP throughput = (TCP Window Size / RTT)
Without packet loss (and retransmissions), this is correct.
TCP throughput = BDP / RTT = (Link Speed in Bytes/sec * RTT)/RTT = Link Speed in Bytes/sec
This is correct only if the window size is configured to the optimal value. BDP/RTT is the optimal (maximum possible) transfer rate in the network. Most modern operating systems should be able to auto-configure it optimally.
How is throughput dependent on Block size ? Is Block size equals TCP window or UDP datagram size?
I don't see any block size in the iperf documentation.
If you refer to the TCP window size, if it is smaller than BDP, then your throughput will be suboptimal (because you waste time waiting for ACKs instead of sending more data; if needed I can explain further). If it is equal or higher to the BDP, then you achieve optimal throughput.
It depends on how you define "Throughput". It usually can be one of the followings.
Number of bytes (or bits) sent in a fixed period of time;
Number of bytes (or bits) sent and received on the receiver end in a fixed period of time;
You can apply these definition to every layer when people talking about throughput. In application layer, 2nd definition means the bytes have really been received by the receiver end of the application. Some people refer to it as "goodput". In Transport layer, say TCP, 2nd definition means the corresponding TCP ACKs are received. To me, most of people should be only interested in the bytes are really received by the receiver end. So, 2nd definition is usually what people mean by "Throughput".
Now, once we have a clear definition of throughput (2nd definition). We can discuss how to measure the throughput correctly.
Usually, people either use TCP or UDP to measure the network throughput.
TCP: People usually measure TCP throughput only on the sender end. As for packets successfully received by the receiver end, ACKs will be sending back. So, sender itself will know how many bytes are sent and received on the receiver end. Divided this number by the measuring time, we will know the throughput.
But, there are two things need to be noticed during TCP throughput measurement:
Is sender side always full buffer during the measurement? i.e. During the measurement period, sender should always has packets to send. It is important for correct throughput measurement. e.g. if I set my measuring time to be 60 seconds, but my file has been finished transmission in 40 seconds. Then there are 20 seconds the network is actually idle. I will under-estimate the throughput.
TCP rate is regulated by its congestion window size, slow-start duration, sender window (and receiver window) size. Sub-optimal configuration of these parameters will result in under-estimated TCP throughput. Although most of the modern TCP implementation should have a quite good configuration of all of these, it is hard for a tester to 100% sure all these configurations are optimal.
Due to these limitations/risks of TCP in network throughput estimation, quite a number of researchers will use UDP for measuring network throughput.
UDP: As UDP has no ACK sending back once the packets are successfully received, people has to measure the throughput in the receiver end. Or, if the receiver end is not easily accessed, people can compare the logs on both sender and receiver sides to determine the throughput. But, this inconvenience is mitigated by some throughput measuring tools. For example, iperf has embedded sequence numbers in its customized payload, so that it can detect any loss. Also, a receiver's report will be sent to the sender to show the throughput.
As UDP by nature is just sending whatever it has to the network and not waiting for the feedback. Its throughput (remember the 2nd definition) once measured will be the actual capacity (or bandwidth) of the network.
So, usually, the throughput measured by UDP should be higher than that from TCP although the difference should be small (~5%-10%).
One biggest drawback of UDP throughput measuring is that, when using UDP one should also make sure that sender-side buffer must be full. (Otherwise, it results in under-estimated throughput as TCP). This step will be little tricky. In iperf, one can specify the sending rate by -b option. Increase -b value in different rounds of testing will converge the throughput measured. For example, in my gigabit ethernet, I first use -b 100k in the test. The throughput measured is 100Kbps. Then I perform the following iterations to converge the maximum throughput which is the capacity of my ethernet.
-b 1m --> throughput: 1Mbps
-b 10m --> throughput: 10Mbps
-b 100m --> throughput: 100Mbps
-b 200m --> throughput: 170Mbps
-b 180m --> throughput: 175Mbps (this should be quite close to the actual capacity)
From what I studied, congestion avoidance phase sets CWND = CWND + MSS * (MSS/CWND) every time a new acknowledgment is received. This is assuming we don't encounter duplicate ACKS or timeouts. But what happens if there are delayed acknowledgements ?
Here's what I think from research on delayed acks (no idea if this is correct):
Basically Delayed ACK is the destination retaining the ACK segment for a period of time expecting one of two things.
Either there will be more ACKS will be required to be sent before the timer is up because of new packets recieved by the receiver. OR the receiver will need to send some data back to the sender in which case it can piggy back the message on that packet.
How does this affect the congestion avoidance phase ?
This would be bad for congestion avoidance phase of TCP which depends on new Acks to increase CWND. This would cause delays in CWND window size change thus causing delay in the sending of packets. This means by the time that TCP could be sending packets to the receiver, it is actually not because acknowledgments are being delayed.
This affects the congestion avoidance phase the same way it affects the other phase (SS) : it will slow down the traffic. However, keep in mind there are two different network uses, the interactive one (such as telnet), and the bulk one. Delayed Acks are likely to be used with interactive protocols sending very small amounts of data, but this can bring new problems if Nagle's algorithm is used by the other side. When unsure, just disable delayed Acks.
That is a really good question. Since they keep the bottleneck buffer full, delayed Ack is not a big problem for traditional congestion control algorithms such as Reno and CUBIC.
For TCP Vegas which tries to keep the queue small, it is still not a problem because if it faces delayed Ack, it will reduce cwnd very slowly (one unit every RTT). Therefore a Vegas connection will not suffer from under-utilization unless delayed ack lasts for a very long time.
TCP slow start came about in a time when the Internet began experiencing "congestion collapses". The anecdotal example from Van Jacobson and Michael Karels paper goes:
During this period, the data throughput from LBL to UC Berkeley (sites separated
by 400 yards and two IMP hops) dropped from 32 Kbps to 40 bps.
The congestion problem is often described as being caused by the transition from a high-speed link to a slow-speed link, and packet build up/dropping at the buffer at this bottleneck.
What I'm trying to understand is how such a build up would cause a drop in end-to-end throughput, as opposed to simply causing superfluous activity/retransmits on the high-speed portion of the link leading into the full buffer. As an example, consider the following network:
fast slow fast
A ======== B -------- C ======== D
A and D are the endpoints and B and C are the packet buffers at a transition from a high speed to low speed network. So e.g. the link between A/B and C/D is 10Mbps, and link between B/C is 56Kbps. Now if A transmits a large (let's say theoretically infinite) message to D, what I'm trying to understand is why it would take it any longer to get through if it just hammered the TCP connection with data versus adapting to the slower link speed in the middle of the connection. I'm envisaging B as just being some thing whose buffer drains at a fixed rate of 56Kbps, regardless of how heavily its buffer is being hammered by A, and regardless of how many packets it has to discard because of a full buffer. So if A is always keeping B's buffer full (or overfull as may be the case), and B is always transmitting at it's maximum rate of 56Kbps, how would the throughput get any better by using slow-start instead?
The only thing I could think of was if the same packets D had already received were having to be retransmitted over the slow B/C link under congestion, and this was blocking new packets. But wouldn't D have typically ACK'd any packets it had received, so retransmitted packets should be mostly those which legitimately hadn't been received by D because they were dropped at B's buffer?
Remember that networks involve sharing resources between multiple computers. Very simplistically, slow start is required to avoid router buffer exhaustion by a small number of TCP sessions (in your diagram, this is most likely at points B and C)
From RFC 2001, Section 1:
Old TCPs would start a connection with the sender injecting multiple
segments into the network, up to the window size advertised by the
receiver. While this is OK when the two hosts are on the same LAN,
if there are routers and slower links between the sender and the
receiver, problems can arise. Some intermediate router must queue
the packets, and it's possible for that router to run out of space.
[2] shows how this naive approach can reduce the throughput of a TCP
connection drastically.
...
[2] V. Jacobson, "Congestion Avoidance and Control," Computer
Communication Review, vol. 18, no. 4, pp. 314-329, Aug. 1988.
ftp://ftp.ee.lbl.gov/papers/congavoid.ps.Z.
Routers must have finite buffers. The larger a speed mismatch between links is, the greater the chance of buffer exhaustion without slow start. After you have buffer exhaustion, your average TCP throughput will go down because buffering increases TCP's ability to utilize links (preventing unnecessary drops for instantaneous link saturation).
Note that RFC 2001 above has been superseded by RFC 5681; however, RFC 2001 offers a more quotable answer to your question.
From your OP...
Now if A transmits a large (let's say theoretically infinite) message to D, what I'm trying to understand is why it would take it any longer to get through if it just hammered the TCP connection with data versus adapting to the slower link speed in the middle of the connection.
First, there is no such thing as an infinite message in TCP. TCP was limited by the initial window size before slow-start came along.
So, let's say the initial TCP segment was 64KB long. If the entire TCP segment fills the router's tx buffer at B, TCP utilizes less of the link over time due to dynamics involved with packet loss, ACKs and TCP back-off. Let's look at individual situations:
B's tx_buffer < 64KB: You automatically lost time for retransmissions because A's TCP is sending faster than B can dequeue packets
B's tx_buffer >= 64KB: As long as A is the only station transmitting, no negative effects (as long as D is ACK-ing correctly); however, if there are multiple hosts transmitting on A's LAN trying to transit across the 56K link, there are probably problems because it takes 200 milliseconds to dequeue a single 1500 byte packet at 56K. If you have 44 1500-byte packets from A's 64KB initial window (44*1460=64KB; you only get 1460 bytes of TCP payload), the router has a saturated link for almost 9 seconds handling A's traffic.
The second situation is neither fair nor wise. TCP backs off when it sees any packet loss... multiple hosts sharing a single link must use slow start to keep the situation sane.
BTW, I have never seen a router with 9 seconds of buffering on an interface. No user would tolerate that kind of latency. Most routers have about 1-2 seconds max, and that was years ago at T-1 speeds. For a number of reasons, buffers today are even smaller.
While talking about UDP I saw/heard congestion come up a few times. What does that mean?
congestion is when you are trying to send too much data over a limited bandwidth, it cannot send the data faster than the incoming amount so additional packets are dropped.
When congestion occurs, you can see these effects:
Delay due to the queue at one end of the connection being too big, so it takes time for your packet to be transmitted.
Packet loss when new packets are simply dropped, forcing connection resets (and often causing more congestion).
Lower quality of service, protocols like TCP will do a cutback on the transmission rate, so your throughput will be lowered.
Blocking, certain networks have protocol priorities, so your UDP packets may be dropped in favor of allowing TCP traffic through.
Its like a traffic jam, imagine right after a sports game where a parking lot full of cars is trying to empty out into a small side street.
It means that network-connected devices are attempting to send more data across the network than it can handle, e.g. 20 Mbps of data across a 10 Mbps link.
In context of UDP, it's your main source of lost datagrams under ordinary circumstances.
Most LANs use some sort of a collission detection/avoidance system. A congestion typically means that the amount of data which is being transmiited on the medium is causing enough collissions to deteriorate the quality of service defined for that medium.
You may want to read up CSMA/CD at wikipedia.
As UDP packets can often be broadcasted, congestion can occur more often.
Kind regards,
For instance, Ethernet is a broadband protocol. Once a message is sent, every node receives it but ignores if the packet are not sent to them. What happens when two nodes send a packet at the same time? It will cause a collision and data loss.
So, both of the nodes will have to resend the message. To avoid more collisions, nodes are designed to wait a random number of milliseconds. Otherwise they keep going on sending messages simultaneously and packages will collide forever.