AWS re:Invent 2022 -- Keynote with Peter DeSantis
-
-
Peter本次keynote的主题是Performance!
我们来看看这次时长接近1个小时15分钟的keynote里,各个新AWS的invent是如何在performance这个主线上被穿针引线的。
本系列合集“#天锥re:Invent2022视点”assume读者已对AWS的产品、组件、特性,以及其他术语有基本的了解。故这里不对基础概念做过多介绍,而是专注于这次re:invent的内容本身。
AWS Nitro v5 (GA)
新版Nitro带来的性能改善:
-
60% Higher PPS
-
30% Lower packet latency
-
40% Better performance per Watt(等效功耗的降低)
C7gn EC2 instance
C7gn搭载了新版Nitro v5芯片。截至re:invent 2022,C7gn还处在Preview阶段。C7gn的性能提升:
-
200Gbps Networking bandwidth
-
Up to 50% better packet processing performance, compared to C6gn
-
2x Higher PPS/vCPU, compared to C5gn
Graviton3E CPU
相比Graviton3,Graviton3E在vector instruction处理方面有35%的性能提升。
HPC7g EC2 instance
截至re:invent,HPC7g 的状态是coming soon。不知到了2023年初,那时是否可以一睹它的身影。
HPC7g搭载了Nitro v5和Graviton3E芯片。
EFA (Elastic Fabric Adapter)这个不是新东西了。主要和HPC、ML场景有关。
SRD (Scalable, Reliable, Datagram) - AWS研发的持续低延时网络协议。Multipath这个本质上不是什么新鲜的东西。运行在Nitro上,AWS专属。性能改善:
-
Lower latency
-
High throughput
SRD的应用场景一:EBS
-
90% lower EBS tail latency
-
4x EBS throughput with SRD
扩展阅读:
-
https://assets.amazon.science/a6/34/41496f64421faafa1cbe301c007c/a-cloud-optimized-transport-protocol-for-elastic-and-scalable-hpc.pdf
-
https://aws.amazon.com/blogs/hpc/in-the-search-for-performance-theres-more-than-one-way-to-build-a-network/
SRD的应用场景二:ENA (Elastic Network Adapter) Express (GA),性能改善:
-
44% lower tail latency
-
25GBps (TCP over ENA Express) single connection throughput,对TCP over ENA的5GBps
ElastiCache支持ENA express (coming soon),尤其对于high-9s tail latency的降低。
EC2上启用ENA Express: https://aws.amazon.com/blogs/aws/new-ena-express-improved-network-latency-and-per-flow-performance-on-ec2/
Trn1n EC2 instance(coming soon)
Serverless computing
-
Lambda SnapStart - reduce Lambda function cold start time
-
Built on Firecraker snapshot functionality
-
Restoring Uniqueness in MicroVM Snapshots (paper) - https://arxiv.org/pdf/2102.12892.pdf
keynote的最后,Peter分享了本次re:invent2022与performance话题有关的session:https://reinvent.awsevents.com/performance-at-reinvent/
纵观这个keynote,与我而言,最让人感兴趣的就是Lambda SnapStart这个新特性。后面会安排专题对其进行deep dive。
-