Layer 4 Loadbalancing for MPTCP

Layer 4 loadbalancing is widely used in datacenters. In order to ensure the smooth operation of a high-capacity web service, incoming requests from clients must be evenly spread across a large number of servers that service said requests. Datacenter operators place a number of loadbalancers between the border routers and the servers. The loadbalancers are tasked with ensuring that all packets belonging to a connection reach the same server. This is typically achieved via a hashing-based scheme: a hash function is applied to each packet's 5-tuple and a destination server is chosen based on the result. Multipath TCP is an extension to TCP designed to take advantage of the multiple paths that may exist between two hosts by using an arbitrary number of subflows. Current loadbalancer designs do not work with Multipath TCP, and this has hampered the adoption of the protocol, especially on the server side. The main reason why loadbalancing MPTCP connections is not trivial is because a connection's individual subflows look like independent TCP connections. There is no discernible relationship between their 5-tuples, and loadbalancing them based on a hash of the 5-tuple will most likely result in them reaching different servers. Furthermore, connection-identifying information can only be extracted from the initial 3-way handshake of each flow. This document proposes two mutually-exclusive solutions to these problems. They rely to varying degrees on getting the client to embed connection or server-identifying information in the packets that it sends out. This extra information can be used statelessly by the loadbalancers. Both solutions require modifications only to the server stack and work well with existing MPTCP clients.

Our first proposal revolves around controlling the destination port that the client uses in all subflows aside from the initial one. It is possible for the server to advertise an additional port via the ADD_ADDR option . This informs the client that it can send an MP_JOIN to this new port and initiate a new subflow. To take advantage of this, each server is be assigned a unique 16-bit ID, which must be different from the port on which the service is being hosted (e.g. 80). As soon as a connection is initiated, the server sends an ADD_ADDR to the client advertising a new port equal to said ID. Packets that arrive at the loadbalancer are treated as follows: Packets destined to the port that the service is being hosted on will be forwarded to a server based on a hash of the 5-tuple. Packets destined to any other port are forwarded to the server whose ID matches the destination port. This approach has two drawbacks: The client will most likely also try to initiate subflows using the server's original port. Because these subflows are loadbalanced based on a hash of their 5-tuple, they will almost certainly reach a different server and break. (Using REMOVE_ADDR to prevent the creation of these subflows would entail the destruction of the original subflow.) If the client is behind a firewall that restricts access to certain destination ports, it might not succeed in establishing any new subflows.

Our second proposal is to loadbalance packets based on the server's token. The token's most significant 14 bits are treated as a hash value for the connection. They are embedded in all outgoing TCP timestamps, and subsequently echoed back by the client. Incoming packets that do not contain timestamps (such as FINs) are dealt with via redirection between the servers.

The client initiates an MPTCP connection by sending a SYN with the MP_CAPABLE option. Under normal operation, the server then picks a random 64-bit key for the connection, and uses it to compute its token. To forward the packet appropriately, the loadbalancer must know the token before deciding what server to send it to. To accomplish this, we move the key generation to the loadbalancer. The connection's token can be computed based on the generated key. The loadbalancer places the generated key, along with the IP of the server that would be responsible for the subflow under normal 5-tuple hashing (which we call the alternate server IP) in an IP option and forwards the SYN to the server.

The figure above depicts the IP option that is inserted into the MP_CAPABLE packet before it is sent to the server. We have chosen an IP option despite the fact that the data contained therein pertains to the transport layer, because TCP option space is very limited. IP option type 96 is currently classified as reserved . Upon receipt of the packet, the server uses the key provided to compute the token for the connection. If no connection with the same token exists, the server uses the key provided. Otherwise, it takes a brute-force approach and randomly generates multiple keys and selects one that yields a token with the same 14 highest-order bits. The use of the alternate server IP will be discussed in a later section.

Additional subflows are initiated by the client by sending MP_JOIN packets. These packets contain the server's token. Similarly to how MP_CAPABLE packets are treated, the loadbalancer uses an IP option to inform the server about which other server would be responsible for the subflow under normal 5-tuple hashing.

IP option type 97 is also classified as reserved .

The TCP timestamp option is present in most packets and is comprised of two fields: the TSval, which is set by the packet's sender, and TSecr, which contains a timestamp recently received from the other end. Taking advantage of the fact that timestamps set by the server are echoed back by the client, the server shifts its timestamp clock left by 14 bits, and embeds the 14 highest-order bits of the token into the 14 lowest-order bits of the TSval. When a packet with the ACK flag set and with the TS option present arrives at the loadbalancer, it is forwarded based on the 14 least significant bits of the TSecr field.

Timestamps supplied by the server are used by the client for protection against wrapped sequence numbers (PAWS). We assume that the server uses a timestamp clock frequency of 1 tick per ms, which is the highest frequency recommended by . The recycling time of the timestamp clock's sign bit is required to be greater than the Maximum Segment Lifetime of 255 seconds. Given that the clock ticks once every ms in increments of 2 ^ 14, its recycling time is roughly 262 s, which is within the bounds set by the standard. While the quickly-increasing timestamp is benign to active subflows, PAWS will still cause segments to be dropped if the subflow in question had been idle for a period longer than the clock's recycling time. To solve this, the server periodically sends keepalive messages during idle periods.

Some packets (most notably FINs) do not contain timestamps or any other connection-identifying information. As such, they are forwarded to a server based on a hash of the 5-tuple. As seen in and , whenever a new subflow is setup, the server responsible for it (A) also knows which other server (B) would be hit by the packets in case 5-tuple hashing is used. A will use a simple peer-to-peer protocol to inform B to setup a redirection rule for the 5-tuple in question. The redirection rule will be deleted by B either at A's request, after the subflow has finished, or after a timeout. We do not discuss the specifics of the protocol in this document. Redirection of a packet is performed using IP-in-IP encapsulation.

The ability to perform layer 4 loadbalancing in a scalable manner is crucial for the adoption of Multipath TCP. This document explored two ways in which this can be accomplished. We put forth that loadbalancing is feasible with the current version of MPTCP and that significant changes to the protocol for loadbalancing support are unnecessary.