restart (minimal caching) • Warp test to measure network saturation when using S3 • Impala TPCDS benchmark • Ratis streaming performance tests How much does disk read cost with NVME? Impala TPCDS Why datanode instead of the Leader ● Netty zero buffer copy ○ No gRPC buffer problem Ratis streaming Benchmark – Observation ● V2 Streaming multiple-client cases can be ~3x of V1 Async ○ Streaming can use

所有已存在的桶，升级后变成LEGACY 版本，以⽀持向后兼容 存储Key格式基本同OBS, 通过配置项区分偏向⽂件，还是偏向S3对象的⽀持 在Bucket 创建时指定Layout，后续不⽀持更改 17 Benchmark Query Details: Dropped “catelog_sales” table with sub- paths(files/dirs) count = 5K Query Completion ~= 16 million 数据块 ⾜够! • 16TB HDD, 极⼩⽂件场景，数据块平均1KB, 总共 16 billion 数据块 有挑战! 21 Micro Benchmark Container 数量上升 I. V2 性能明显下降 II. V3 性能波动不⼤(除Container删除操作） Single node(96C 250G), 10 disks(3TB)

restart (minimal caching) • Warp test to measure network saturation when using S3 • Impala TPCDS benchmark • Ratis streaming performance tests Software under test CDP Private Cloud Base 7.1.8 + • IMPALA-11457