Notes on performance hierarchies
The hardware components between interconnection levels can also affect bandwidth, as follows:
A drive has sequential access bandwidth and average latency times for seek and rotational delays.
Drives perform optimally when doing sequential I/O to disk. Non-sequential I/O forces movement of the disk head (that is, seek and rotational latency). This movement is a huge overhead compared to the amount of data transferred. The more non-sequential I/O done, the slower it becomes.
Simultaneously reading or writing more than one stream results in a mix of short bursts of sequential I/O with seek and rotational latency in between. This situation significantly degrades overall throughput, especially when the drive approaches 85% capacity. Different drive types have different seek and rotational latency specifications. Therefore, the type of drive has a large effect on the amount of degradation.
From best to worst, such disk drives are Serial Attach SCSI (SAS). "SATA" Serial ATA drives are rarely used for enterprise backup systems. SATA Solid-State Disks (SSDs) have access times at least twice that of their rotating disk counterparts and a 0.1 versus 12-millisecond access time. NVMe Gen4 SSDs have speeds of 3.8Gb/s writes and 7GB/s reads. For large repositories of data that is kept available, but if speed of retrieval is not a prime concern disk drive is the largest most cost efficient storage medium.
A RAID controller has cache memory of varying sizes. The controller also does the parity calculations for RAID-5 or RAID-6. RAID-5 can be used on SSD systems as the rebuild time is fast enough to not have a concurrent failure. RAID-6 or RAID-10 is required for Disk-based storage.
A PCIe card can be limited either by the speed of the ports or the clock rate to the PCIe slot. More information about the size and speed of the PCIe configurations is available:
Memory can be a limit if there is intensive non-I/O activity in the system. Ensure that you follow the 1GB of memory to 1TB of MSDP storage so that the SHA-2 encryption is efficiently handled.
While CPU performance contributes to all performance, it is not the bottleneck in most modern systems for I/O intensive workloads. Very little work is done at that level. The CPU must execute a read operation and a write operation, but those operations do not take up much bandwidth. An exception fingerprint calculation in MSDP deduplication is CPU intensive. High number of concurrent streams with a high deduplication ratio can create a CPU bottleneck. Plan for at least one CPU thread per backup stream on lower powered systems. Up to four concurrent streams per core (two threads per core) on Intel processors are possible with small and numerous file backups