Saturday, November 15, 2014

Paper Summary: Granola, Low overhead distributed transaction coordination

This paper is by Cowling and Liskov, and it appeared at Usenix ATC'12.  Most closely related papers to this paper are the Sinfonia and Calvin papers. So, it may be helpful to also read the summaries of those papers from the links above to familiarize yourself with them.

This paper looks at coordination of 1-round transactions, which are different from general transactions that involve multi-round interaction with the client. 1-round transactions execute at the participant nodes, with no communication with other nodes except for at most a single commit/abort vote.

Figure 1 shows the system architecture. The clients are the transaction managers for these 1-round transactions. They initiate transactions and evaluate the commit/conflict/abort decisions returned for the transactions. The repositories communicate with one another (for at most a single commit/abort vote) to coordinate transactions.

There are 2 types of transactions in Granola: coordinated ones, and uncoordinated ones. Actually, better/less-confusing names for these 2 types would be lock-coordinated transactions and timestamp-coordinated transactions, since the latter type of transactions still involve coordination among the repositories.

To reflect this duality in coordination mode, each repository runs in one of two modes. When there are no coordinated transactions running at a repository, it runs in timestamp mode. When a repository receives a request for a coordinated transaction, it switches to the locking mode.

Uncoordinated transactions

There are two kinds of uncoordinated transactions: single repository transactions, or independent distributed transactions. Granola uses a timestamped-based coordination to provide serializability to both kinds of uncoordinated transactions and avoids locking for them. Examples of independent distributed transactions include read-only transactions, and local-read transactions, such as give every employee a 2% raise.

(In our previous work, the slow-fast paper (2010), we had also made an analysis similar to the independent transactions idea. Our slow-fast analysis inspects the program actions, which are precondition guarded assignment statements, and determined for which actions the atomicity can be relaxed so that they can execute in an uncoordinated manner. Our finding was that if the precondition of a program action is "locally-stable" (i.e., this precondition predicate cannot get falsified by execution of other program actions), then it is safe to execute this program action in an uncoordinated/nonatomic manner.)

Repositories are assumed to have access to loosely synchronized clocks (say NTP synchronized clocks). The use of timestamps in Granola are for ordering/serializability of transactions, and logical clocks would also suffice. Granola needs time synchronization for performance/throughput not for correctness. This separation is always a welcome one. (I can't resist but refer to our hybrid logical clock, HLC, work here. I think HLC would improve exposition of Granola, and would make it tolerant to malformed clients which may increase highTS and deny service to normal clients' requests.)


Figure 5 shows the straightforward single repository execution, and Figure 6 shows the more interesting independent transaction execution. Since these are 1-round transactions, the commit/abort decisions are local and one-shot at repositories. Voting is used to notify other repositories about the local decision (a conflict vote cause a repository to abort the transaction), and also for nominating a proposed timestamp for the transaction. The transaction is assigned the highest timestamp from among the votes.

In the "pure" timestamp mode, all single repository and independent distributed transactions commit successfully, and serializability is guaranteed as they are executed in timestamp order. This provides a substantial reduction in overhead from locking and aborts, which improve the throughput.

While in the timestamp mode, coordinated transactions may also arrive which may lead to conflict decisions to be returned for the single or distributed independent transactions executing on those repositories. Those repositories will switch to coordinated mode to serve the coordinated transactions. Once all coordinated transactions have completed those repositories can again transition to the timestamp mode.

Dually, in locking mode, repositories can still serve single repository and distributed independent transactions provided they do not conflict, i.e., they do not need to update a locked item. (In effect, the timestamp mode is just a shortcut to denote the lack of any coordinated transactions in the system.)

Coordinated transactions

Figure 7 shows coordinated transaction execution. Coordination is needed for a transaction that requires a remote read for abort/commit decision. For example, the transaction to transfer $50 from Alice's account to Bob's account first needs to remote-read Alice's account to certify that it indeed has more than $50. (Recall that in contrast for an independent distributed transaction the read, or the guard of the transaction, was local: give every employe 2% raise.) Different from independent distributed transaction execution, the coordinated transaction execution involves a prepare phase to acquire required locks. Locks ensure serializability for coordinated transaction execution. So timestamp order may not be satisfied in commit of transactions, and thus, external consistency may not be provided for coordinated transaction execution.

Discussion

This paper is written with a focus on describing the system in reasonable detail to potential users of the system. This diverges a bit from the academic style, which would put the focus on the novelties and "intellectual merit" (a la NSF). Maybe this is because of the style/focus of the USENIX ATC conference. The evaluation section is also really nice and detailed. (Granola performs like Sinfonia for coordinated transacations, and improve throughput for uncoordinated transactions.)

It seems that Granola does not offer much advantage over Calvin. The paper compares in the related work with Calvin and states the following. "The Calvin transaction coordination protocol was developed in parallel with Granola, and provides similar functionality. Rather than using a distributed timestamp voting scheme to determine execution order, Calvin delays read/write transactions and runs a global agreement protocol to produce a deterministic locking order."

I think Granola may have an edge over Calvin for low-latency, especially for transactions that involve a couple repositories. The best way to see why is to consider this analogy.
Granola:Calvin :: CSMA:TDMA.

No comments: