Why do we need an LDBC SNB-specific workload driver?
In a previous 3-part blog series we touched upon the difficulties of executing the LDBC SNB Interactive (SNB) workload, while achieving good performance and scalability. What we didn't discuss is why these difficulties were unique to SNB, and what aspects of the way we perform workload execution are scientific contributions - novel solutions to previously unsolved problems. This post will highlight the differences between SNB and more traditional database benchmark workloads. Additionally, it will motivate why we chose to develop a new workload driver as part of this work, rather than using existing tooling that was developed in other database benchmarking efforts. To briefly recap, the task of the driver is to run a transactional database benchmark against large synthetic graph datasets - "graph" is the word that best captures the novelty and difficulty of this work.
Workload Execution - Traditional vs Graph
Transactional graph workloads differ from traditional relational workloads in several fundamental ways, one of them being the complex dependencies that exist between queries of a graph workload.
To understand what is meant by "traditional relational workloads", take the classical TPC-C benchmark as an example. In TPC-C Remote Terminal Emulators (emulators) are used to issue update transactions in parallel, where the transactions issued by these emulators do not depend on one another. Note, "dependency" is used here in the context of scheduling, i.e., one query is dependent on another if it can not start until the other completes. For example, a New-Order transaction does not depend on other orders from this or other users. Naturally, the results of Stock-Level transactions depend on the items that were previously sold, but in TPC-C it is not an emulator's responsibility to enforce any such ordering. The scheduling strategy employed by TPC-C is tailored to the scenario where transactional updates do not depend on one another. In reality, one would expect to also have scheduling dependencies between transactions, e.g., checking the status of the order should only be done after the order is registered in the system. TPC-C, however, does not do this and instead only asks for the status of the last order for a given user. Furthermore, adding such dependencies to TPC-C would make scheduling only slightly more elaborate. Indeed, the Load Tester (LT) would need to make sure a New-Order transaction always precedes the read requests that check its status, but because users (and their orders) are partitioned across LTs, and orders belong to a particular user, this scheduling does not require inter-LT communication.
A significantly more difficult scheduling problem arises when we consider the SNB benchmark that models a real-world social network. Its domain includes users that form a social friendship graph and which leave posts/comments/likes on each others walls (forums). The update transactions are generated (exported as a log) by the data generator, with assigned timestamps, e.g. user 123 added post 456 to forum 789 at time T. Suppose we partition this workload by user, such that each driver gets all the updates (friendship requests, posts, comments and likes on other user's posts etc) initiated by a given user. Now, if the benchmark is to resemble a real-world social network, the update operations represent a highly connected (and dependent) network: a user should not create comments before she joins the network, a friendship request can not be sent to a non-existent user, a comment can only be added to a post that already exists, etc. Given a user partitioning scheme, most such dependencies would cross the boundaries between driver threads/processes, because the correct execution of update operations requires that the social network is in a particular state, and that state depends on the progress of other threads/processes.
Such scheduling dependencies in the SNB workload essentially replicate the underlying graph-like shape of its dataset. That is, every time a user comments on a friend's wall, for example, there is a dependency between two operations that is captured by an edge of the social graph. Partitioning the workload among the LTs therefore becomes equivalent to graph partitioning, a known hard problem.
Because it's a graph
In short, unlike previous database benchmarking efforts, the SNB workload has necessitated a redefining of the state-of-the-art in workload execution. It is no longer sufficient to rely solely on workload partitioning to safely capture inter-query dependencies in complex database benchmark workloads. The graph-centric nature of SNB introduces new challenges, and novel mechanisms had to be developed to overcome these challenges. To the best of our knowledge, the LDBC SNB Interactive benchmark is the first benchmark that requires a non-trivial partitioning of the workload, among the benchmark drivers. In the context of workload execution, our contribution is therefore the principled design of a driver that executes dependent update operations in a performant and scalable way, across parallel/distributed LTs, while providing repeatable, vendor-independent execution of the benchmark.