Throughput of storage systems

For data backup, the data from one or more storage systems is read and written to tape; for data recovery, data is read from tape and written to disk.

Storage systems are connected to the server via Fibre Channel using multiple paths or via SAS (JX). This permits full use to be made of their performance capability.

If the options of parallelization (suitable RAID groups, PAV, multiplexing) are used, the degree of connection utilization can be extremely high in the case of throughput-oriented data backup loads.

Performance recommendations

• Use of state-of-the-art, high-speed storage systems with FBA disks D3435 and data format NK2. These disks can be operated with an IO length of up to 480 KB.

• Use of storage systems with as large a cache as possible. The size selected should ensure that the read hit rates in OLTP mode are 60 - 80%.

• Multipath connection of the storage systems to the server.

• Use of RAID level RAID 1/0, RAID 5, or RAID 6 (see section "Replication: volume-based mirroring for storage systems"). This also enables “large” volumes (more than 32 GB in size) to be used.
  Fewer parallel IOs can be executed with RAID 1. RAID 1 is therefore less suitable when a high throughput is to be achieved.

• Use of dynamic PAV on /390 servers, see "Advantages of PAV".

Characteristics of the data backup load

For data backup, very large quantities of data are read and written sequentially from and to the storage system in several parallel, asynchronous data streams. Current storage systems can can read from the physical disks using read-ahead mechanisms and write to the physical disks using Delayed Fast Writes. The limits of the caches can be reach when doing this. Consequently, an appropriate cache size must be ensured.

The performance of the physical disks and of the RAID groups is, however, also decisive for the throughput which can be achieved. With a data backup load, today's storage systems achieve a very good throughput rate with read-ahead or Delayed Fast Write.

Measurements

To determine the maximum throughput of backup volumes of the storage system, only reading from the storage system to BS2000 main memory is performed using a test program based on ARCHIVE without writing to tape. The storage system has a multipath connection.

The table below shows the possible throughput with different settings. The measured values reflect the ideal case with a small number of large files (1000 MB). They also show the reduction in throughput when medium-sized (10 MB) or small (1 MB) files are used instead of the optimum very large size.

The throughput rates specified refer to one task during the actual backup phase, in other words without starting and ending the programs, and without creating the report files.

The measurement results in this section assume favorable conditions, e.g. no parallel load on the measured configuration, sufficient CPU performance.

SE700 server, ETERNUS DX600-S3 storage system, RAID 1/0, connection via Fibre Channel (8 Gbit/s, 2-path)

File size	Read throughput (Mbyte/s) without PAV	Read throughput (Mbyte/s) with PAV (1 alias)
1000 MB	263	388
10 MB	250	314
1 MB	147	202

The throughput loss compared to large files is low in the case of medium file sizes (10 MB). In the case of small files (approx.1 MB), it becomes significant.

Measurements on an SE300 server on a comparable storage system yield the same results as on an SE700 with 1 PAV alias.

For data backup, "data striping" (RAID 1/0, RAID 5, RAID 6) provides RAID groups with a higher throughput than RAID 1.