Example implementations disclosed herein can be used to build, maintain, and access databases built database in multi-core computing systems with large VRAM and huge NVRAM. The database with optimistic concurrency control can be built on a transactional key-value data store that includes logically equivalent data pages stored in both VRAM and VRAM. Data records in volatile data pages in the VRAM represent the most recent version of the data. Data records in the NVRAM immutable and are organized in a stratified composite snapshot. A distributed log gleaner process is used to process log entries corresponding to transactions on the volatile data pages and construct the snapshot. The log gleaner sorts the log entries by epoch, key range, and most recent use to partition the snapshot across multiple nodes.

Networkswith Remote DirectMemoryAccess (RDMA) support are becoming increasingly common. RDMA, however, offers a limited programming interface to remote memory that consists of read, write and atomic operations. With RDMA alone, completing the most basic operations on remote data structures often requires multiple round-trips over the network. Data-intensive systems strongly desire higher-level communication abstractions that supportmore complex interaction patterns. A natural candidate to consider is MPI, the de facto standard for developing high-performance applications in the HPC community. This paper critically evaluates the communication primitives of MPI and shows that using MPI in the context of a data processing system comes with its own set of insurmountable challenges. Based on this analysis, we propose a new communication abstraction named RDMO, or Remote DirectMemory Operation, that dispatches a short sequence of reads, writes and atomic operations to remote memory and executes them in a single round-trip.

In example implementations, a method include receiving a request for a lock in a Mellor-Crummey Scott (MCS) lock protocol from a guest user that is context free (e.g., a process that does not bring a queue node). The lock determines that it contains a null value. The lock is granted to the guest user. A pi value is received from the guest user to store in the lock. The pi value notifies subsequent users that the guest user has the lock.

Techniques are provided to allow more sophisticated operations to be performed remotely by machines that are not fully functional. Operations that can be performed reliably by a machine that has experienced a hardware and/or software error are referred to herein as Remote Direct Memory Operations or “RDMOs”. Unlike RDMAs, which typically involve trivially simple operations such as the retrieval of a single value from the memory of a remote machine, RDMOs may be arbitrarily complex. The techniques described herein can help applications run without interruption when there are software faults or glitches on a remote system with which they interact.

RDMA has been an important building block for many highperformance distributed key-value stores in recent prior work. To sustain millions of operations per second per node, many of these works use hashing schemes, such as cuckoo hashing, that guarantee that an existing key can be found in a small, constant number of RDMA operations. In this paper, we explore whether linear probing is a compelling design alternative to cuckoo-based hashing schemes. Instead of reading a fixed number of bytes per RDMA request, this paper introduces a mathematical model that picks the optimal read size by balancing the cost of performing an RDMA operation with the probability of encountering a probe sequence of a certain length. The model can further leverage optimization hints about the location of clusters of keys, which commonly occur in skewed key distributions. We extensively evaluate the performance of linear probing with a variable read size in a modern cluster to understand the trade-offs between cuckoo hashing and linear probing. We find that cuckoo hashing outperforms linear probing only in very highly loaded hash tables (load factors at 90% or higher) that would be prohibitively expensive to maintain in practice. Overall, linear probing with variable-sized reads using our model has up to 2.8× higher throughput than cuckoo hashing, with throughput as high as 50M lookup operations per second per node.

Hybrid concurrency control in a database transaction is disclosed. A contention value associated with a database object is read. The contention value is dynamically updated based on access patterns of the database object. If the contention value indicates the database object is under high contention, a pessimistic concurrency control mechanism is applied to the database object. An optimistic concurrency control mechanism is applied to the database object if the contention value does not indicate the database object is under high contention.

A technique includes performing graph inference in a graph inference engine that includes multiple processing nodes to determine assignments for vertices of a graph. Performing the graph inference includes controlling remote memory accesses within the engine, including storing first data in a local memory of the first processing node, where the first data represents at least assignments for a plurality of vertices of the graph; in the first processing node, determining updates for the assignments for a subset of the plurality of vertices of a partition of the graph assigned to the first processing node and modifying the first data based on the updates; and communicating the updates to at least one other processing node of the multiple processing nodes, where at least one other partition of the graph is assigned to the other processing node(s).

Example implementations disclosed herein include techniques for methods, systems, and devices for a foster twin data structure. The data structure includes data pages, keys, and pointers. The pointers define hierarchical parent-child relationships among the data pages based on the keys. Transactions that result in inserts into the data pages are handled using foster twin data pages that split key ranges and the corresponding data records into two foster twin data pages. New pointers are added to the parent data page of the split data page to indicate the new parent child relationship between the parent data page and the foster twin data pages. The old data page can be retired by adding a moved-bit and removing the old pointers from the parent data page to guide concurrent transaction and prevent future transactions.

Example implementations disclosed herein include techniques for a ready only bufferpool for use in local nodes of a multi-node computing system. Read only transactions executed by a processor can reference a ready only bufferpool resident in a VRAM on the same node. If the desired data page is in the bufferpool, the transaction can immediately read data records from the cached data pages. If the desired data page is not in the bufferpool, then the transaction can cause a copy of a corresponding data page in a secondary memory to be installed in the bufferpool. The bufferpool can include more than one copy of a data page simultaneously to handle and prevent cache line misses. Data page are dropped from the bufferpool based on an incrementing per data page counter.

Distributed query processing is often performed by a set of nodes that apply MapReduce to a data set and materialize partial results to storage, which are then aggregated to produce the query result. However, this architecture requires a preconfigured set of database nodes; can only fulfill queries that utilize MapReduce processing; and may be slowed down by materializing partial results to storage. Instead, distributed query processing can be achieved by choosing a node for various portions of the query, and generating customized code for the node that only performs the query portion that is allocated to the node. The node executes the code to perform the query portion, and rather than materializing partial results to storage, streams intermediate query results to a next selected node in the distributed query. Nodes selection may be involve matching the details of the query portion with the characteristics and capabilities of the available nodes.

Existing scale-up graph engines are tuned for either short, navigational requests (e.g., Nearest-Neighbor) or longer, analytics requests (e.g., PageRank). However, they do not have good performance for both workloads running concurrently. We present Janus, a scale-up graph engine archi-tected for modern, many-core servers with large memory. Janus has excellent scale-up performance on navigational requests, on analytics requests, and on a mixed workload running concurrently both navigational and analytics requests

Example implementations disclosed herein include techniques for systems, methods, and devices for a heap data structure organized into linked-lists of epoch data pages on a per-core basis in a multi-core multi-node computing system to handle many concurrent transactions

The MCS lock is one of the most prevalent queuing locks. It provides fair scheduling and high performance on massively parallel systems. However, the MCS lock mandates a bring-your-own-context policy: each lock user must provide an additional context (i.e., a queue node) to interact with the lock. This paper proposes MCSg, a variant of the MCS lock that relaxes this restriction. Our key observation is that not all lock users are created equal . We analyzed how locks are used in massively-parallel modern systems, such as NUMA-aware operating systems and databases. We found that such systems often have a small number of "regular" code paths that enter the lock very frequently. Such code paths are the primary beneficiary of the high scalability of MCS locks. However, there are also many "guest" code paths that infrequently enter the lock and do not need the same degree of fairness to access the lock (e.g., background tasks that only run periodically with lower priority). These guest users, which are typically spread out in various modules of the software, prefer context-free locks, such as ticket locks. MCSg provides these guests a context-free interface while regular users still enjoy the benefits provided by MCS. It can also be used as a drop-in replacement of MCS for more advanced locks, such as cohort locking. We also propose MCSg++, an extended version of MCSg, which avoids guest starvation and non-FIFO behaviors that might happen with MCSg. Our evaluation using microbenchmarks and the TPC-C database benchmark on a 16-socket, 240-core server shows that both MCSg and MCSg++ preserve the benefits of MCS for regular users while providing a context-free interface for guests.

In one example, a unified lock includes a first level lock and a second level lock. The first and second level locks allow threads executing on a processor to request exclusive control of a processor resource. A first thread in the second level lock acquires an individual lock in the first level lock and then passes the individual lock to successive threads in the second level lock within a bounded limit before relinquishing the individual lock to remaining threads in the first level lock.

A computing system includes a graph engine comprising an embedded database. The graph engine may: store vertices of a graph in the embedded database as key-value pairs associated with each of the vertices. Each of the key-value pairs may comprise information associated the vertices. The graph engine may further: receive a request indicating operations to perform on a plurality of the vertices, partition the graph into partitions comprising of a subset of the vertices of the graph, partition the vertices into blocks comprising a subset of the vertices of the partitions, and perform the indicated operations on the plurality of the vertices of the request.

Future servers will be equipped with thousands of CPU cores and deep memory hierarchies. Traditional concurrency control (CC) schemes---both optimistic and pessimistic---slow down orders of magnitude in such environments for highly contended workloads. Optimistic CC (OCC) scales the best for workloads with few conflicts, but suffers from clobbered reads for high conflict workloads. Although pessimistic locking can protect reads, it floods cache-coherence backbones in deep memory hierarchies and can also cause numerous deadlock aborts. This paper proposes a new CC scheme, mostly-optimistic concurrency control (MOCC), to address these problems. MOCC achieves orders of magnitude higher performance for dynamic workloads on modern servers. The key objective of MOCC is to avoid clobbered reads for high conflict workloads, without any centralized mechanisms or heavyweight interthread communication. To satisfy such needs, we devise a native, cancellable reader-writer spinlock and a serializable protocol that can acquire, release and re-acquire locks in any order without expensive interthread communication. For low conflict workloads, MOCC maintains OCC's high performance without taking read locks. Our experiments with high conflict YCSB workloads on a 288-core server reveal that MOCC performs 8× and 23× faster than OCC and pessimistic locking, respectively. It achieves 17 million TPS for TPC-C and more than 110 million TPS for YCSB without conflicts, 170× faster than pessimistic methods.