Connection management without retaining information

Abstract

. AbstractManaging a connection between two hosts in a network is an important service to provide in order to make the network useful for many applications. The two main subproblems are the management of serial incarnations of a connection and the transfer of messages within an incarnation. This paper investigates whether it is necessary for connection management protocols to retain state information across node crashes and between incarnations. The following results were obtained:When information is not retained across node crashes, incarnation management i s not possible at all. When information is not retained between incarnations, incarnation management i s possible if the network is FIFO and not possible if the network is non-FIFO. When information is not retained across node crashes, message transfer can be accomplished in networks that lose packets if the network is FIFO and the protocol is allowe d a v ariable length grace period after a crash during which it need not deliver messages. However, message transfer cannot be accomplished if the network is non-FIFO or the grace period allowed is xed. When information is not retained across node crashes, message transfer can be accomplished in networks that do not lose packets if the network is FIFO or the protocol need not be FIFO. Message transfer is not possible when the network is non-FIFO and the protocol must be FIFO. If the network has bounded capacity, then message transfer is possible without using stable storage. This indicates, somewhat surprisingly, that there is a data link initialization protocol that can withstand node crashes without stable storage.

Exon: An Oblivious Exactly-Once Messaging Protocol

“…Nevertheless, EO guarantees in TCP only hold within a connection/session; when the connection fails (likely to occur in current WAN environments and long-lived communications), it either allows for message loss or duplication, as Belsnes [6] shows for any single-message communication. Attiya et al [2] proved that when state information is not saved between incarnations, the problem is solvable if and only if the network is FIFO-which is not the case for most networks. Therefore, to ensure EO, it is necessary to retain inter-connection information, e.g., at an upper layer.…”

Scalable eventually consistent counters over unreliable networks

“…By the induction hypothesis, S would have existed until T was acquired, and as S is absent from C, this implies that T was already acquired. (2) Or T could be removed by the tokens map entry for (s, d) being overwritten in createtoken; this last case is not possible because: tokens are created using the current source clock, which is then incremented; for a token with id (s, d, sck s , ) to be created, a received counter state C from d must contain a slot (s, d, sck , ) with sck = sck s . This means that d would have previously received T from s and from the induction hypothesis, d would have had a corresponding slot and would have already acquired T filling the slot.…”

“…If an infinite duration connection incarnation is maintained for each client, where no operation can fail due to a timeout, this will imply stable server state to manage each client connection, leading to state explosion in servers, as clients cannot be forgotten. This because there is no protocol that gives reliable message transfer within an infinite incarnation for general unbounded capacity (e.g., wide-area networks) non-FIFO lossy networks that does not need stable storage between crashes [2,9]. This problem can be overcome by never failing due to a timeout, but allowing connections to be closed if there are no pending requests and the connection close handshake is performed successfully (e.g., before a partition occurs).…”