What speed should you be seeing for backups?

In order to understand the speed of NetBackup, you need to understand the hardware that you have...

Tape Drives
ò DDS2 rated as being able to write 1.6GB/hour native and 3.2GB/hour compressed.
ò DDS3 3.5GB/7GB per hour
ò DDS4 10GB/20GB per hour
ò DLT4000 5GB/10GB per hour
ò DLT7000 18GB/36GB per hour
ò DLT8000 21GB/42GB per hour
ò SDLT 38GB/76GB per hour
ò LTO 53GB/106GB per hour
ò LTO (Gen2) 100GB/200GB per hour
To calculate what you expect to be the data rate for backups, take the following calculation:
Tape Drive MB/Sec x(Compression factor û Usually 1.5) x 3600 / 1024 to give you your rate per hour in GigaBytes. e.g. A DLT7000 rated at 5MB/sec is:
5MB/sec x 1.5 x 3600 / 1024 = 26.4GB/hr. A LTO will be 79GB/hr.

NICÆs
ò 10Mb/sec is 1.25MB/sec (8 bits in a byte) which translates to 4.4GB/hour transfer (IEEE 802.3i)
ò 100Mb/sec is 44GB/hour (IEEE 802.3u)
ò 1000Mb/sec is 440GB/hour (IEEE 802.3ab which is Ether and not fibre)
ò TCP/IP has overhead so the actual transfer is typically around 50-60% efficiency. If your NICÆs support it, set the checksum offload to Tx/Rx so that the NIC handles the CRC calculations instead of the CPU. This does reduce CPU overhead and improves backup speed marginally.
ò Set the correct Max Transfer Unit (MTU) size û To calculate ôping server -f -l <Packet Size-28k for IP and ICMP headers). When size is exceeded the packet will get broken up.
ò To calculate the best Receive Window size (RWIN), calculate your MTU and then ping the server and insert the results into:
Bandwidth(KBPS) / 8 * Average Latency(MiliSec) = RWIN Size(Bytes). Round the result up so that it is exactly divisible by your Maximum Segment Size (MSS). If MTU is 1500 then MSS is 1460 (40k overhead).

Duplexing
ò Half-Duplex is akin to a walkie talkie û Only one side can talk at a time
ò Full-Duplex is analogous to a telephone conversation so both sides can talk at a time.
The debate on this issue is never ending. From testing etc the best configuration that we have found is to set both NIC and switch to Full-Duplex and the Max Speed such as 100Mb. On auto-negotiate we have found there to be many problems and backup speeds drop off considerably when network connections time-out û This is applicable to NICÆs rated as 10/100. On GB NICÆs, use auto-negotiate as it automatically implies full-duplex.

Flow-Control
Another area that there does not seem to be much agreement on in the IT world but from our testing we have found that if we set the NICÆs and the switch to enable flow control for Rx and Tx, we get considerably better performance when there is a lot of data flowing to the media servers. This is especially important in a switched environment. In testing and tracing the TCP/IP packets what we found was that the media server quite often resends packets because it thinks that they were lost but the are getting queued on the switch and when the network is busy, the packets do actually get dropped, also resulting in a resend.

SCSI Cards
It is important to understand how much data flows through a SCSI channel so that you do not connect too many devices to any one channel.
ò SCSI 5MB/sec (8 bit bus)
ò SCSI-2 (Fast SCSI) 10MB/sec (8 bit bus)
ò SCSI-2 (Fast Wide SCSI) 20MB/sec (16 bit bus)
ò Ultra SCSI 20Mb/sec (8 bit bus)
ò SCSI-3 (Ultra-Wide SCSI) 40MB/sec (16 bit bus)
ò Ultra 2SCSI (LVD) 40Mb/sec (8 bit bus)
ò Ultra Wide2 SCSI (LVD) 80MB/sec (16 bit bus)
ò Ultra Wide3 SCSI (LVD) 160MB/sec (16 bit bus)
ò Fibre Channel SCSI 100MB/sec
For example, an Adaptec 2944UWD card can transfer 140GB/hour and an Adaptec 39160 can transfer 563GB/hour per channel so 1126GB/hour total.

Bus Speed
ò PCI runs at 33MHz so a 64bit 33MHz bus is rated at 264MB/sec.
ò A 32bit 33MHz bus is rated at 132MB/sec.
ò As a comparison an 8x AGP video card running in a 2x mode can transfer 2133MB/sec.

Fragmentation
Fragmentation of hard drives can result in a 10-20% performance loss, depending on the level of fragmentation. With a GB Ethernet connection to a client from a media server that has a DLT7000 drive attached you should be getting in the range of about 25GB/hour. With fragmentation this may drop to around 20GB/hour. As an example, take a server that has 1000 users opening 50 files a day at a delay of 0.5sec/file û Just server performance is a loss of 6.94 hours a day.

OTM (Now Called VSP)
To me, this is a dirty word. If you do not absolutely need it, do not use it. In a test where we had 2GB in 25,000 files we got 1600KB/sec transfer rate using OTM. Turning OTM off increased it to 5300KB/sec.

CPU and Memory
For each 1MB/sec of data arriving at the server you need 5MHz CPU processing power. To calculate:
<Num of clients> * <rate in MB/sec> * 5MHz * 2 (5MHz for reading off network and 5MHz for writing to tape). E.g. 8 jobs averaging 6000KB/sec needs = 480MHz.
Memory Used = (buffer_size * num_buffers) * num_drives * MPXÆing.