Today, we are releasing Seastar, a new open-source C++ framework for extreme high-performance applications on OSv and Linux. Seastar brings a 5x throughput improvement to web-scale workloads, at millions of transactions per second on a single server, and is optimized for modern physical and virtual hardware.
Benchmark results are available from the new Seastar project site.
Today’s server hardware is substantially different from the machines for which today’s server software was written. Multi-core design and complex caching now require us to make new assumptions to get good performance. And today’s more complex workloads, where many microservices interact to fulfil a single user request, are driving down the latencies required at all layers of the stack. On new hardware, the performance of standard workloads depends more on locking and coordination across cores than on performance of an individual core. And the full-featured network stack of a conventional OS can also use a majority of a server’s CPU cycles.
Seastar reaches linear scalability, as a function of core count, by taking a shard-per-core approach. Seastar tasks do not depend on synchronous data exchange with other cores which is usually implemented by compare-exchange and similar locking schemes. Instead, each core owns its resources (RAM, NIC queue, CPU) and exchanges async messages with remote cores. Seastar includes its own user-space network stack, which runs on top of Data Plane Development Kit (DPDK). All network communications can take place without system calls, and no data copying ever occurs. Seastar is event-driven and supports writing non-blocking, asynchronous server code in a straightforward manner that facilitates debugging and reasoning about performance.
Seastar is currently focused on high-throughput, low-latency network applications. For example, it is useful for NoSQL servers, for data caches such as memcached, and for high-performance HTTP serving. Seastar is available today, under the Apache license version 2.0.
Please follow @ScyllaDB on Twitter for updates.