InfoSlice's Performance. TechnoMages, Inc.

Flexibility | Reliability | Performance | User-friendly | Security

Disk Array Performance

TechnoMages, Inc.

Performance factors

The performance of any RAID depends on a number of factors, including the RAID type (e.g., striped with parity; mirror; striped mirror-pairs, etc.), the number of disk access arms in the RAID set, the Operating System of the test machine (storage client), the chip set and drivers of the HBA and NIC (or, for SCSI and Ethernet, perhaps board-level ports), and the type of benchmark program.

Sustainable throughput of different storage interface types

To reveal the relative efficiencies, in general, of Parallel SCSI and Fibre Channel HBAs and their drivers, of draft8 iSCSI client drivers and Gigabit Ethernet chip sets and drivers, and the target drivers and parity stripe processing (RAID 5) of the TechnoMages InfoSlice, we measured the sustained streaming read and write throughputs of 64KB requests for both Windows and Linux servers. The number of disk drives used in each case was sufficient to saturate the burst rate of the storage interface; and the performance-enhancing effect of caches along the path was removed by setting the overall transfer amount to a size at least 30x the size of the largest cache.

Windows 2000 Server iometer benchmark
Interface	Reads (64KB requests) MB/s *	Writes (64KB) MB/s *
2 Gbps Fibre Channel	178	151
1Gbps Fibre Channel	93	95
Ultra 160 Parallel SCSI	104	93
1Gbps Ethernet, iSCSI draft8	81	55

* A note on the units of rate measurements:

Megabytes reported by disk I/O tests such as iometer are 2**20 bytes (1,048,576), rather than 10**6 bytes, which is important because the signal rates of 1Gb/s Fibre Channel and Gigabit Ethernet use the 10**9 meaning of Giga-, and this meaning usually gets carried to the advertised MB/s capacity of these links.

Thus, for example, after the overhead fields, interframe idles, acknowledgment frames (which need not always be sent but for this calculation we take the worse case of one ACK frame per payload-frame), and the 8B/10B encoding of Fibre Channel are subtracted from the nominal line rate of 1.0625 Gbps, the bandwidth available for Fibre Channel payloads (such as SCSI data blocks) is 100.369 M_[base10]Bps. If no ACK frames at all were to be sent, the value would be 103.225 M_[base10]Bps. (For streaming I/O, the available bandwidth is likely closer to the larger number.)
To appreciate the values that iometer reports, then, the above Mega_[base10] values for can be converted to their Mega_[base2] equivalents, which are 95.72 M_[base2]Bps and 98.44 M_[base2]Bps.

Linux lmbench/lmdd benchmark ‡
Interface	Reads (64KB requests) MB/s	Writes (64KB) MB/s
1Gbps Fibre Channel	86‡	79
Ultra 160 Parallel SCSI	95	77
1Gbps Ethernet, iSCSI draft8	55	34

‡ The lmbench/lmdd benchmark does not take advantage of disk request queues, thus the lower values in this table relative to Windows-iometer, do not reflect OS differences per se. We plan to publish Linux-iometer measurements, also.