Feature List

x86-TaiShan hybrid deployment:

• In block/object/file storage services, x86 and TaiShan servers are deployed in the same storage pool.
• In block/object/file storage services, TaiShan servers are used to expand the capacity of an x86 server cluster.

The same Ceph software version and OS software version are used.

In block/object/file storage services, NVMe SSDs are used as the cache of HDDs to improve system performance.

The OS page size needs to be changed from 64 KB to 4 KB.

I/O pass-through improves the storage performance in balanced configuration and 7:3 read/write hybrid scenarios.

It cannot improve the storage performance when the bcache is configured.

• The TaiShan 200 server (model 2280) has passed the certification by commercial-edition Ceph for Ubuntu.
• The TaiShan 200 server (model 5280) has passed the certification by commercial-edition Ceph for Ubuntu.

The Kunpeng BoostKit for SDS compression algorithm is a Huawei proprietary lossless compression algorithm. Compared with open source compression algorithms, it delivers a higher compression rate and performance.

Constraint

It is invalid for incompressible data sources such as images and videos.

Supported on VMs: No

Based on service workload characteristics, data is prefetched to the cache to increase the read cache hit ratio and the IOPS performance in read operations. In sequential read operations, the IOPS performance is increased by 20% for I/Os of less than 512 KB.

1. The bcache with kernel version 4.14 is used.

2. Block storage service is used.

3. The service workload is sequential read.

The EC Turbo feature of Kunpeng BoostKit for SDS optimizes the erasure code (EC) process of the open source Ceph and decreases the I/O amplification ratio in the data read/write process.

1. The block storage or object storage service is used.

2. Mixed read/write (7:3).

3. Ceph 14.2.8 is used.

4. The bcache feature is not supported.

The smart write cache consists of the bcache kernel patch and related tools. The smart write cache uses I/O passthrough, Bcache QoS policy control, write back policy control, and GC policy control to improve the bcache performance and the Ceph cluster performance ultimately.

1. CentOS 7.6, kernel 4.14.

2. openEuler 20.03 LTS SP1, kernel 4.19.

3. On CentOS, the smart write cache and smart I/O prefetch cannot be used at the same time, while on openEuler, the smart write cache and smart I/O prefetch can be used at the same time.

4. Storage engine: BlueStore.

The Kunpeng Storage Acceleration Library (KSAL) is a Huawei-developed storage algorithm library that uses algorithms optimized based on Kunpeng instead of mainstream open source algorithms to improve storage performance.

Constraint

Currently, the EC algorithm supports only the 2+2, 4+2, 6+2, 12+3, and 20+3 ratios.

Supported on VMs: No

Metadata acceleration is a storage engine performance acceleration feature developed by Huawei and optimized based on RocksDB.

RocksDB is a high-performance, persistent, and embedded key-value storage engine developed by Facebook. It is widely used in large-scale data storage and processing, such as Internet services, distributed systems, and data analysis services. Based on RocksDB, the metadata acceleration feature uses a Huawei-developed algorithm to enable Kunpeng acceleration for better storage performance. This feature fits well with the Kunpeng architecture to optimize read and write hotspots, adjust background tasks (data flushing and compaction) based on service loads, and optimize cache logic based on data hotspots.

The Ucache smart read cache uses smart I/O prefetch to accurately identify hotspot requests, prefetch I/Os of the sequential pattern, interval pattern and more, and load I/Os to the read cache in advance. In addition, the read cache uses the LRU algorithm to evict cold data, improving the I/O hit ratio and read performance.

• The maximum capacity transferred during cache initialization is 256 TiB. The value cannot be dynamically changed.
• The value of cache_line_size can be 8 KiB, 16 KiB, 32 KiB, or 64 KiB. 8 KiB is recommended.
• ocf creates io_worker_num queues. One io_worker corresponds to one ocf queue pair (submission_queue/completion_queue).
• region_id of each region is globally unique. One slot corresponds to one core. A maximum of 511 cores are available. The maximum logical space of a core is 4,096 TiB, and can host a maximum of 128,000 32 GiB regions. All regions in a slot are placed in the core corresponding to the slot. In the device space, region_id is remapped to a remap_id, and the region range on the core is remap_id x 32 GiB to (remap_id + 1) x 32 GiB.
• The ocf_get, ocf_put, ocf_invalid, and ocf_lookup interfaces called by the same slot must run in the same thread.

The data compaction algorithm of the Kunpeng BoostKit for SDS is deployed on the open-source distributed storage cluster Ceph to eliminate data waste caused by zero padding. In addition, combined with functions including data encapsulation, space allocation based on block counting, granularity-based traffic diversion, batch submission, and batch callback, the data compaction algorithm improves the data reduction ratio and overall system IOPS, which reduces costs and improves performance.

1. Ceph 14.2.8 is used.

2. The block or object storage service is used.

The Kunpeng BoostKit storage maintenance tool library provides the HDD/SSD fault prediction and slow disk detection algorithm library to predict storage system media exceptions in advance and improve storage system stability.

HDD/SSD fault prediction is based on machine learning algorithms. Smart data is collected to train models to predict and identify potential faulty disks in a storage cluster. Predicts disk faults before they affect services. This helps customers to handle disk faults in a timely manner and prevents service losses caused by disk faults.

BoostIO uses memory and disk resources on the computing side to build distributed multi-level cache. Write cache improves service write performance and data reliability. The read cache increases the read cache hit ratio, improving service read performance.

3. The cache medium specifications of BoostO are memory and NVMe SSD disks. That is, each compute node must be configured with memory and NVMe SSD disks as the data cache medium space for BoostIO.

A plugin is applied to the Ceph network framework AsyncMessage to support Unified Communication X (UCX), which enables full RDMA in Ceph all-flash storage.