Watching OSDI 2020 presentation videos

I set watching at least one OSDI'20 presentation video per day during the long vacation as one of my plans.

A large scale analysis of hundreds of in-memory cache clusters at Twitter

https://www.youtube.com/watch?v=OQtMM5vdhlI&feature=emb_title

The authors traced and analyzed the workloads on the Twitter’s in-memory cache systems. To me, below findings were interesting.

• There are many write requests.
• Size of each object is not so big (median 200 bytes), so metadata, which sizes 64 bytes per each object, is a burden
• Size of each key is not small compared to the size of each value.

The trace data is available via github[1]. It’s 2.8TB for compressed version, and 14TB for uncompressed version.

[1] https://github.com/twitter/cache-trace

AIFM: High-Performance, Application-Integrated Far Memory

https://www.youtube.com/watch?v=xHhaniGXTUg&feature=emb_title

OS-level fast-network-based far memory approaches provide transparency to the application but wastes performance due to the semantic gap due to the fixed-size page abstraction and the kernel space time consumption. AIFM solves the semantic gap using new data structure abstraction and provides user space runtime that don’t need kernel space time consumption. As a result, the application is required to be modified, but the authors argue it’s only modest change. Compared to other state-of-the-art (FastSwap from EuroSys'20), it achieved 13x speedup.

LinnOS: Predictability on Unpredictable Flash Storage with a Light Neural Network

https://www.youtube.com/watch?v=yzv9lcjxhAg&feature=emb_title

It’s not easy to predict each I/O speed of modern fast storage devices, due to their complex internals like the caching, the wear-leveling and the garbage collection. One well-known solution is hedging. It prepare an array of SSDs, issue I/O to one of them, and if the response doesn’t made until a timeout, revoke the request and try with another SSD. The wait time limit bounds the latency. LinnOS uses an approach similar to the hedging, but it uses a neural network that can predict if each I/O to each SSD will be served fast or not. For this, the neural network receives current I/O queue depth and queue depths and latencies of last few I/Os as input. Then, it predicts if the latency will be only fast or slow. To mitigate with the effect from wrong predictions, it uses biased learning and adaptive hedging based on the prediction accuracy. The network is learned offline.

Do OS abstractions make sense on FPGAs?

https://www.youtube.com/watch?v=k-cp4U3JKug&feature=emb_title

Computer systems embedding FPGA in addition to CPU, called heterogeneous computing systems, are widesparead nowadays. These systems typically have high operational complexity. Especially developing and deploying FPGA application is quite complicated. For the reason, some manufacturers provide some FPGA shells providing some of the OS abstractions. The authors developed a FPGA shell called Coyote, which provides full abstraction sets of the OS and experimented if it works well for the systems. The result was very positive.

Fast RDMA-based Ordered Key-Value Store using Remote Learned Cache

https://www.youtube.com/watch?v=Qv-0YL_SII4&feature=emb_title

Because server-centric key-value store (KVS) performance bound to the server CPUs, RDMA-based approaches were proposed. Because one RTT is required for one RDMA, the approaches require many round trips and therefore doesn’t show high performance. Caching the index in client is one solution, but it makes huge client memory footage. The authors propose to use neural network as the index cache. They train neural network to receive key and provide logical address of the value for the key in server. The network retrained for dynamci updates and copyied to clients. Then, client use the nerual network, which is much smaller than the full index tree to know the address of the value and fetch it via the RDMA.

A Simpler and Faster NIC Driver Model for Network Functions

https://www.youtube.com/watch?v=zKJIY4vbBDA&feature=emb_title

Software-defined network approach, which implements functionality of appliances like bridege, router, firewall is widespread. For better performance and flexibility, the software complexity has increased so that it’s hard to do formal verification of the network stacks. The authors argue that by sacrificing some of the fliexibility, it’s available to implement simple and fast network stack. They prove their idea with a new network driver model, tinynf. They implemented a driver for Intel 82599 based on the driver model with only 550 lines of code. It was able to finish formal verification 7x faster than a state-of-the-art driver. Also, it achieved 1.6x performance compared to the state-of-the-art.

Theseus: an experiment in operating system structure and state management

https://www.youtube.com/watch?v=i1pLDZKtlBI

In this paper, the authors present an OS called Theseus that designed for state spill problem. For this, they made it to composed with many tiny components called ‘cell’, and applied rust-like language level safe guaranteeness mechanisms in the OS level.

Specification and verification in the field: Applying formal methods to BPF just-in-time compilers in the Linux kernel

https://www.youtube.com/watch?v=2V3ts5-W_9g&feature=emb_title

BPF program is verified before run, but it runs after compiled via JIT, after the verification. Therefore, if there is a bug in the JIT, real problem occurs. The authors of this paper devloped a new BPF JIT for RISC-V with a JIT correctness specification framework, jitterbug. The changes made to the kernel are merged into the mainline kernel.

Storage Systems are Distributed Systems (So Verify Them That Way!)

https://www.youtube.com/watch?v=4s8EeXcu_8Y&feature=emb_title

Storage systems have high complexity, so it’s hard to do the verification. The authors of this paper realized the architecture and charactersitics of the storage systems are similar to those of distributed systems, and applied the verification methodologies for distributed systems to the storage systems after making it more general. Based on this, they implemented a verifiable key-value storage, VeriSafeKV. The performance of it was similar to unverified DB systems.

Caladan: Mitigating Interference at Microsecond Timescales

https://www.youtube.com/watch?v=G-v3ndwixOI&feature=emb_title

On a system, multiple workloads that has different requirements co-scheduled. For example, some workloads only need best-effort resources while others require strict tail latency. Because the workloads share some hardware resources like LLC and memory bandwidth, it’s hard to fulfill all the requirements. Well know solution is resource partitioning. Because static partitioning could result in low resource utilization, dynamic partitioning solutions were previously proposed. However, the dynamic partitioning doesn’t provide microsecond granularity decision, while 100 microseconds is the marginal timewindow to guarantee the tail latency problem, according to the authors’ arguments. The authors propose to use different interference signals and design the system to work in micro-second level.

Semeru: A Memory-Disaggregated Managed Runtime

https://www.youtube.com/watch?v=MFA3MmNDKaM&feature=emb_title

Constructing system as distributed machines for different works such as CPU works and memory works are know ans resource-disaggregated architecture and gaining popularity. However, the architecture is usually written for native applications rather than GC-based applications. This paper introduces JVM runtime, Semeru, which is designed for resource-disaggregation architecture.

PANIC: A High-Performance Programmable NIC for Multi-tenant Networks

https://www.youtube.com/watch?v=EB6dK3L8Jzg&feature=emb_title

Multi-tenant networks require generality, flexible chaining, isolation, and performance to programmable NIC. However, current programmable NICs doesn’t support those. Authors propose a new programmable NIC design for the requirements and introduce an implementation on FPGA, PANIC.

Serving DNNs like Clockwork: Performance Predictability from the Bottom Up

https://www.youtube.com/watch?v=wHOpY_MY57Y&feature=emb_title

Nowadays, data center serves for many machine learning models for different users. Even though the latency of each inference is predictable because it has no conditional branches, the models serving system could result in unpredictable end-to-end latency. This paper proposes a new machine learning models serving system that designed do protect the latency.

Teting Configuration Changes in Context to Prevent Production Failures

https://www.youtube.com/watch?v=QrGKmp3ALKQ&feature=emb_title

In this paper, a new test framework called ‘ctest’ is proposed for validation of configuration changes that can made in production environment. It’s similar to other test frameworks like Junit in its form.