- May 28, 2004
- TCP Tuning Domains
- TCP State Model
- TCP Congestion Control and Flow Control Sliding Windows
- TCP and RDMA Future Data Center Transport Protocols
- About the Author
TCP Congestion Control and Flow Control Sliding Windows
One of the main principles for congestion control is avoidance. TCP tries to detect signs of congestion before it happens, and reduce or increase the load into the network accordingly. The alternative of waiting for congestion and then reacting is much worse because once a network saturates, it does so at an exponential growth rate and reduces overall throughput enormously. It takes a long time for the queues to drain, and then all senders again repeat this cycle. By taking a proactive congestion avoidance approach, the pipe is kept as full as possible without the danger of network saturation. The key is for the sender to understand the state of the network and client and to control the amount of traffic injected into the system. Flow control is accomplished by the receiver sending back a window to the sender. The size of this window, called the receive window, tells the sender how much data to send. Often, when the client is saturated it might not be able to send back a receive window to the sender, signaling it to slow down transmission. However, the sliding windows protocol is designed to let the sender know, before reaching a meltdown, to start slowing down transmission by a steadily decreasing window size. At the same time these flow control windows are going back and forth, the speed at which ACKs come back from the receiver to the sender provides additional information to the sender which caps the amount of data to send to the client. This is computed indirectly.
The amount of data that is to be sent to the remote peer on a specific connection is controlled by two concurrent mechanisms:
The congestion in the network - The degree of network congestion is inferred by the calculation of changes in Round Trip Time (RTT): that is the amount of delay attributed the network. This is measured by computing how long it takes a packet to go from sender to receiver and back to the client. This figure is actually calculated using a running smoothing algorithm due the large variances in time. The RTT value is an important value to determine the congestion window, which is used to control the amount of data sent out to the remote client. This provides information to the sender on how much traffic should be sent to this particular connection based on network congestion.
Client load - The rate at which the client can receive and process incoming traffic. The client sends a receive window that provides information to the sender on how much traffic should be sent to this connection, based on client load.
TCP Tuning for ACK Control
FIGURE 10 shows how senders and receivers control ACK waiting and generation. The general strategy is that clients want to reduce receiving many small packets. Receivers try to buffer up a bunch of received packets before sending back an acknowledgment (ACK) to the sender, which will trigger the sender to send more packets. The hope is that the sender will also buffer up more packets to send in one large chunk rather than many small chunks. The problem with small chunks is that the efficiency ratio or useful link ratio utilization is reduced. For example, a one-byte data packet requires 40 bytes of IP and TCP header information and 48 bytes of Ethernet header information. The ratio works out to be 1/(88+1) = 1.1 percent utilization. When a 1500-byte packet is sent, however, the utilization can be 1500/(88+1500) = 94.6 percent. Now, consider many flows on the same Ethernet segment. If all flows are small packets, the overall throughput is low. Hence, any effort to bias the transmissions towards larger chunks, without incurring excessive delays is a good thing, especially interactive traffic such as Telnet.
FIGURE 10 TCP Tuning for ACK Control
FIGURE 10 provides an overview of the various TCP parameters. For a complete detailed description of the tunable parameters and recommended sizes, refer to your product documentation or the Solaris AnswerBooks at docs.sun.com.
There are two mechanisms senders and receivers use to control performance:
Senders timeouts waiting for ACK. This class of tunable parameters controls various aspects of how long to wait for the receiver to send back an ACK of the data that was sent. If tuned too short, then excessive retransmissions occur. If tuned too long, then excess wasted idle time elapses before the sender realizes the packet was lost and retransmits.
Receivers timeouts and number of bytes received before sending an ACK to sender. This class of tunable parameters allows the receiver to control the rate at which the sender sends data. The receiver does not want to send an ACK for every packet received because the sender will send many small packets, increasing the ratio of overhead to actual useful data ratio and reducing the efficiency of the transmission. However, if the receiver waits to long, there is excess latency which increases the burstiness of the communication. The receiver side can control ACKs with two overlapping mechanisms, based on timers and the number of bytes received.
TCP Example Tuning Scenarios
The following sections describe example scenarios where TCP require tuning, depending on the characteristics of the underlying physical media.
Tuning TCP for Optical Networks WANS
Typically, WANS are high-speed, long-haul network segments. These networks introduce some interesting challenges because of their properties. FIGURE 11 shows how the traffic changes as a result of a longer, yet faster link, comparing a normal LAN and an Optical WAN. The line rate has increased, resulting in more packets per unit time, but the delays have also increased from the time a packet leaves the sender to the time it reaches the receiver. This has the strange effect that more packets are now in flight.
FIGURE 11 Comparison Between Normal LAN and WAN Packet Traffic
FIGURE 11 shows a comparison of the number of packets that are in the pipe between a typical LAN of 10 mbps/100 meters with RTT of 71 microseconds, which is what TCP was originally designed for, and an optical WAN, which spans New York to San Francisco at the rate of 1 Gbps with RTT of 100 milliseconds. The bandwidth delay product represents the number of packets that are actually in the network and implies the amount of buffering the network must provide. This also gives some insight to the minimum window size, which we discussed earlier. The fact that the optical WAN has a very large bandwidth delay product as compared to a normal network requires tuning as follows:
The window size must be much larger. The current window size allows for 216 bytes. To achieve larger windows, RFC 1323 was introduced to allow the window size to scale to larger sizes while maintaining backwards compatibility. This is achieved during the initial socket connection, where during the SYN-ACK three-way handshake, window scaling capabilities are exchanged by both sides, and they try to agree on the largest common capabilities. The scaling parameter is an exponent base 2. The maximum scaling factor is 14, hence allowing a maximum window size of 230 bytes. The window scale value is used to shift the window size field value up to a maximum of 1 gigabyte. Like the MSS option, the window scale option should only appear in SYN and SYN-ACK packets during the initial three-way handshake. Tunable parameters include:
tcp_wscale_always: controls who should ask for scaling. If set to zero, the remote side needs to request; otherwise, the receiver should request.
tcp_tstamp_if_wscale: controls adding timestamps to the window scale. This parameter is defined in RFC 1323 and used to track the round-trip delivery time for data in order to detect variations in latency, which impact timeout values. Both ends of the connection must support this option.
During the slow start and retransmissions, the minimum initial window size, which can be as small as one MSS, is too conservative. The send window size grows exponentially, but starting at the minimum is too small for such a large pipe. Tuning in this case requires that the following tunable parameters be adjusted to increase the minimum start window size:
tcp_slow_start_initial: controls the starting window just after the connection is established
tcp_slow_after_idle: controls the starting window after a lengthy period of inactivity on the sender side.
tcp_recv_hiwat and tcp_xmit_hiwat: control the size of the STREAMS queues before STREAMS-based flow control is activated. With more packets in flight, the size of the queues must be increased to handle the larger number of outstanding packets in the system.
Both of these parameters must be manually increased according to the actual WAN characteristics. Delayed ACKs on the receiver side should also be minimized because this will slow the increasing of the window size when the sender is trying to ramp up.
RTT measurements require adjustment less frequently due to the long RTT times, hence interim additional RTT values should be computed. The tunable tcp_rtt_updates parameter is somewhat related. The TCP implementation knows when enough RTT values have been sampled, then this value is cached. tcp_rtt_updates is on by default, but a value of 0 forces it to never be cached, which is the same as the case of not having enough for an accurate estimate of RTT for this particular connection.
FIGURE 12 Tuning Required to Compensate for Optical WAN
Tuning TCP for Slow Links
Wireless and satellite networks have a common problem of a higher bit error rate. One tuning strategy to compensate for the lengthy delays is to increase the send window, sending as much data as possible until the first ACK arrives. This way the link is utilized as much as possible. FIGURE 13 shows how slow links and normal links differ. If the send window is small, then there will be significant dead time between the time the send window sends packets over the link and the time an ACK arrives and the sender can either retransmit or send the next window of packets in the send buffer. But due to the increased error probability, if one byte is not acknowledged by the receiver, the entire buffer must be resent. Hence, there is a trade-off to increase the buffer to increase throughput. But you don't want to increase it so much that if there is an error the performance is degraded by more than was gained due to retransmissions. This is where manual tuning comes in. You'll need to try various settings based on an estimation of the link characteristics. One major improvement in TCP is the selective acknowledgement (SACK), where only the one byte that was not received can be retransmitted, not the entire buffer.
FIGURE 13 Comparison Between Normal LAN and WAN Packet Traffic Long Low Bandwidth Pipe
Another problem introduced in these slow links is that the ACKs play a major role. If ACKs are not received by the sender in a timely manner, the growth of windows is impacted. During initial slow start, and even slow start after an idle, the send window needs to grow exponentially, adjusting to the link speed as quickly as possible for coarser tuning. It then grows linearly after reaching ssthresh, for finer-grained tuning. However, if the ACK is lost, which has a higher probability in these types of links, then the performance throughput is again degraded.
Tuning TCP for slow links includes the following parameters:
tcp_sack_permitted: activates and controls how SACK will be negotiated during the initial three-way handshake:
0 = no sack disabled:
1 = TCP will not initiate a connection with SACK information, but if an incoming connection has the SACK-permitted option, TCP will respond with SACK information.
2 = TCP will both initiate and accept connections with SACK information.
tcp_dupack_fast_retransmit: controls the number of duplicate ACKS received before triggering the fast recovery algorithm. Instead of waiting for lengthy timeouts, fast recovery allows the sender to retransmit certain packets, depending on the number of duplicate ACKs received by the sender from the receiver. Duplicate ACKs are an indication that possibly later packets have been received, but the packet immediately after the ACK might have been corrupted or lost.
TCP SACK is specified in RFC 2018 TCP selective acknowledgement. TCP need not retransmit the entire send buffer, only the missing bytes. Due to the higher cost of retransmission, it is far more efficient to only resend the missing bytes to the receiver.
Like optical WANs, satellite links also require the window scale option to increase the number of packets in flight to achieve higher overall throughput. However, satellite links are more susceptible to bit errors, so too large a window is not a good idea because one bad byte will force a retransmission of one enormous window. TCP SACK is particularly useful in satellite transmissions to avoid this problem because it allows the sender to select which packets to retransmit without requiring an entire window (which contained that one bad byte) for retransmission.
Adjust all timeouts to compensate for long-delay satellite transmissions and possibly longer-distance WANs, the timeout values must be compensated.