DAOS Middleware overview
•Als PPTX, PDF herunterladen•
1 gefällt mir•164 views
DAOS (Distributed Application Object Storage) is a high-performance storage architecture and software stack that delivers scalable object storage capabilities. It uses Intel Optane memory and NVMe SSDs to provide high IOPS, bandwidth, and low latency storage. DAOS supports various data models and interfaces like POSIX, HDF5, Spark, and Python. It allows applications to access storage with library calls instead of system calls for high performance.
Empfohlen
Empfohlen
Weitere ähnliche Inhalte
Was ist angesagt?
Was ist angesagt? (20)
Ähnlich wie DAOS Middleware overview
Ähnlich wie DAOS Middleware overview (20)
Mehr von Andrey Kudryavtsev
Mehr von Andrey Kudryavtsev (12)
Kürzlich hochgeladen
Kürzlich hochgeladen (20)
DAOS Middleware overview
- 1. DAOS For Applications Mohamad Chaarawi Extreme Scale Architecture & Development
- 2. Intel CorporationCloud & Enterprise Solutions Group 2 2 DAOS Storage Architecture 2 DAOS Storage Engine Intel® Optane Memory SPDK NVMe Interface Metadata, low-latency I/Os & indexing/query Bulk data 3D-NAND/Optane SSD PMDK Memory Interface HDD AI/Analytics/Simulation Workflow DAOS library POSIX I/O HDF5 Spark… Compute Nodes MPI-I/O Python Libfabric • Low-latency & high-message-rate communications • Native support for RDMA & scalable collective operations • Support for iWarp, RoCE, Infiniband, OPA, Slingshot, … • DAOS library directly linked with the applications • No need for dedicated cores • Low memory/CPU footprint • End-to-end OS bypass • Non-blocking, lockless, snapshot support, … • Fine-grained I/O with media selection strategy • Only application data on SSD to maximize throughput • Small I/Os aggregated in pmem & migrated to SSD in large chunks • Full userspace model with no system calls on I/O path • Built-in storage management infrastructure (control plane) • NFSv4-like ACL Delivers high-IOPS, high-bandwidth and low-latency storage with advanced features in a single tier Storage Nodes
- 3. Intel CorporationCloud & Enterprise Solutions Group 3 Aggregate related datasets into manageable and coherent entities • Distributed consistency & automated recovery • Full Versioning • Simplified data management • Snapshot • Cross-tier Migration • Indexing Storage Containers DAOS Container datadatadatafile dir datadatafile dir datadatadatadatafile dir root Encapsulated POSIX Namespace File-per-process DAOS Container datadatadatadatafile datadatadatadatafile datadatadatadatafile datadatadatadatafile DAOS Container datadatadatadataset group datadatadataset group datadatadatadatadataset group group HDF5 « File » Key-value store Graph DAOS Container valuekey valuekey valuekey valuekey valuekey DAOS Container node node node node node node DAOS Container Columnar Database key key key key Value Value Value Value Value Value Value Value
- 4. Intel CorporationCloud & Enterprise Solutions Group 4 Native support for structured, semi-structured & unstructured data models • Built on top of DCPMM • Unconstrained by POSIX serialization • Custom attributes • Data access time orders of magnitude faster (µs) • Scalable concurrent updates & high IOPS • Non-blocking • Enable in-storage computing DAOS Objects DAOS Storage Engine Open Source Apache 2.0 License Data Model Library/Framework Array KV Store Multi-level KV Store key1 val1 key3 val3 @ @ Application NVMe SSD DAOS key1 val1 root @ key3 val3 @ val2 key2 @@ @ @ val2 con’d val2 key2 @ Application
- 5. Intel CorporationCloud & Enterprise Solutions Group 5 5 DAOS Tools Tool dmg daos Target Administrators Users Lustre Equivalent lctl/mkfs/mount/IML lfs Functionality • Storage provisioning • Burn-in • Firmware update • Data plane mgmt & monitoring • Configure/monitor scrubbing • Pool mgmt • Telemetry • Pool query • Container mgmt • Unified namespace mgmt • Container user attributes • Snapshots • Object debugging • POSIX container configuration
- 6. Intel CorporationCloud & Enterprise Solutions Group 6 6 Application Interface Dataset Mover Capacity Tier PFS, S3, HSM, … DAOS Storage Engine Open Source Apache 2.0 License POSIX I/O HPC APPs HDF5 MPI-IO Python Apache Spark Apache Arrow Analytics/AI APPs TensorFlowSEGY
- 7. Intel CorporationCloud & Enterprise Solutions Group 7 DAOS API is new and very flexible • Multi-level keys • Different value types supported • Can build all data models / IO middleware on top of it Most applications still based on POSIX • Need a smooth migration path with little to no application changes • Quick path to realize performance of DCPMM & DAOS POSIX implemented as a middleware instead of being the building block of all data models. POSIX I/O Support Application / Framework DAOS library (libdaos) DAOS File System (libdfs) Interception Library dfuse Single process address space DAOS Storage Engine RPC RDMA End-to-end user space No system calls Intel® QLC 3D Nand SSD
- 8. Intel CorporationCloud & Enterprise Solutions Group 8 MPI-IO Driver for DAOS The DAOS MPI-IO driver is implemented within the I/O library in MPICH (ROMIO). • Added as an ADIO driver • Portable to Open-MPI, Intel MPI, etc. • https://github.com/pmodels/mpich MPI Files use the same DFS mapping to the DAOS Object Model • MPI Files can be accessed through the DFS API • MPI Files can be accessed through regular POSIX with a dfuse mount over the container. Application works seamlessly by just specifying the use of the driver by appending “daos:” to the path. MPI-IO ROMIO driver (https://github.com/pmodels/mpich/tree/master/src/mpi/romio/adio/ad_daos) POSIX / MPI-IO File DAOS Byte Array Object Special DAOS Object: 1 Level Key 1 Byte records Configurable Chunk Size
- 9. Intel CorporationCloud & Enterprise Solutions Group 9The information on this page is subject to the use and disclosure restrictions provided on the second page to this document. HDF5 VOL Architecture Three main components: • HDF5 Library • DAOS VOL Connector • (External) HDF5 Test Suite HDF5 API VOL Layer VFD Layer Native VOL DAOS VOL SEC2 MPIO File System DAOS HDF5 Tools Test Suite New Component … … DAOS APIPOSIX API Core HDF5 Library External Test Suite External VOL Connector Enhanced Component Native Component Through MPI I/O
- 10. Intel CorporationCloud & Enterprise Solutions Group 10 No longer requires separate version of HDF5 • Compatible with main develop branch of HDF5 • Compatible with 1.12.x release series of HDF5 with VOL support Currently supported features • New H5M MAP API to expose K/V interface to HDF5 users • Variable length datatypes are now also supported • Chunking is recommended storage layout to get most of DAOS performance • Tools support (h5dump, h5ls, h5repack etc) Coming by end of the year • Independent metadata writes (= independent object creation) • Asynchronous I/O Available from: https://git.hdfgroup.org/projects/HDF5VOL/repos/daos-vol/ • See user’s guide for more detailed list of supported features HDF5 DAOS VOL Connector – Current Status
- 11. Intel CorporationCloud & Enterprise Solutions Group 11 11 Unified Namespace DAOS Object Store: Addresses pools, containers with uuids; objects with 128-bit IDs. Applications/Users are used to access files / directories in a traditional namespace Unified Namespace allows users to create links between a file/dir in a system namespace to DAOS pools & containers: daos container create --path=/mnt/project1/userA/NS1 --pool=uuid -- type=POSIX/HDF5/etc. Path created above becomes a special file or directory (depending on container type) with an extended attribute with the pool and container information. Accessing that path from DAOS aware middleware will make the link on the fly with the DAOS UNS library.
- 12. Intel CorporationCloud & Enterprise Solutions Group 12 12 Spark Input/Output Support DAOS FileSystem DAOS POSIX API Java Wrapper HDFS Hadoop FileSystem Abstract Class Disks DAOS Storage DAOS Library, DPDK/PMDK/SPDK used, Kernel bypass, zero copy DAOS/Arrow Wrapper Apache Arrow Data Source DAOS API Wrapper DAOS Native Data Source （Parquet, ORC, etc.） Scan/Parse Project, Join, ML, etc. Native Scan/Parse (CSV, Parquet, ORC, etc.) Current Spark Implemented Planned
- 13. Intel CorporationCloud & Enterprise Solutions Group 13 Pythonic bindings called pydaos • Export key-value store objects • Integrated with python dictionaries • Support python iterator, direct assignments, … • Scalable & performant • Bulk insert/retrieve • Core written in C • Python 2.7 & 3 support Python integration • dbm • pyprob TODO • Expose snapshots • Integration with NumPy, … Python Support
- 14. Intel CorporationCloud & Enterprise Solutions Group 14 Resources Source code on GitHub • https://github.com/daos-stack/daos Documentation • http://daos.io Community mailing list • https://daos.groups.io DAOS solution brief • https://www.intel.com/content/www/us/en/high-performance-computing/
- 15. 15
Hinweis der Redaktion
- Overall DAOS architecture Distributed object store built from the ground up for NVM technology. Includes 2 types of storage: SSD with different media, 3d-NAND, Optane. Available with NVMe interface, provide fast block-based storage Intel developing new user-space nvme driver, SPDK Pmem, Optane pmem. Fine grained storage directly memory mapped. Can do direct load store to the storage. No block abstraction Intel developing new storage stack PMDK to manage transactional update to pmem We don’t have any Object store to leverage these object stores. Lustre, GPFS built into kernel. Ceph still uses kernel IOs by going through Linux block layer. Here everything in userspace. DAOS new OS built on top of pmdk and spdk. highIOPS high BW. More advanced storage API, not just byte arrays, KV, etc. Still need to look into communication. Reuse several decade effort to optimize latency and BW for MPI. Re-use same communication middleware as MPI (libfabric). Client side, very lightweight with support to many middleware. All metadata stored on pmem. Small IOs land in pmem and later aggregated into SSD in the background to free space in Pmem. Application large data go in the SSD.
- Baseline API based on Storage container -> Object address space inside a pool. Container can be encapsulated POSIX namespace, or just a flat one. HDF5 has it’s own representations for a single HDF5. KV store can be one. Database, ACG 1 graph Container provides a baseline API with object. Unit of storage management is the container not a file. When you have millions of file, data migration is done on all those. We Simplified that. Fewer number of containers than files, so we can index them. Data migration of entire container. unit of Snapshot. Snapshot not typically available traditionally (used to be at FS level) but now at container level and exposed to user
- POSIX not the foundation any more. New key-array store API provided by DAOS with advanced features (ad-hoc concurrency control, snapshot support, asynchronous API, ..). DAOS API to be directly integrated with rich data model for better performance and extended capabilities. POSIX (i.e. File) support built as a middleware over the DAOS API. DAOS has a native KV API which means that the wire protocol is KV fetch/lookup. On the server, the tree structure is maintained in DCPM and values are either stored in DCPM or SSD depending on sizes. No file serialization is required. A client can directly lookup or insert a key/value pair over the wire and this can happen from many clients concurrently. Keys can be ordered, so we support range query. KV store are partitioned/sharded/replicated/erasure coded over multiple servers for performance & resilience. DAOS can also support complex query since it understand the key-value structure. The native indexing can also be augmented with ad-hoc application-driven index.
- Spark Integration with DAOS – different team at intel Written Hadoop filesystem abstraction class for DAOS. Integrated in DAOS source code. Use DAOS in user-space through libdfs transparently with this Hadoop connector. More value/acceleration to Spark, by integrating DAOS as a datasource through parquet / apache arrow, or through native DAOS API.
- Native Python bindings (pydaos) integrate KV store object directly with python dictionaries. Very pythonic API, you can have iterator, direct assignments, bulk insertion/retrieve Allow people who are not very familiar with C using high level language. Some integration with dbm, pyprob. More work to do