LAN MBytes/Sec for Gbit LAN ?

[Stuart Abramson_2]

Gbit LAN is, of course, rated at 1 Gigabit per second.

But I really care about GBYTES/second, because I think mostly in Bytes. (Like how long will it take to transfer this file..?)

How do I translate Gbit/sec to GByte/sec?
..Do I divide by 8?
..Are there start/stop bits?
..Is there LAN switch overhead?
..Is there an "effective" rate?
..Are there: Rated / Avg / Max values?

[Steven E. Protter]

Divide by 8 will get you theoretical GB/SEc


There is lan switch overhead and your throughput will vary widely based on what kind of switch you have.

In our environment which has Cisco Gigabit core switch, between HP-9000 boxes on large file transfers we'eve seen transfer rates with scp max out at around 9 MB/s

I'm sure that can be topped with a different Core switch environment.

rp5450 2 GB RAM, HP Gigabit LAN card.

SEP

[Michael Tully]

Well start somewhere that convert these:

http://www.frii.com/support/knowledgebase/viewArticle.php?articleID=236

[Hein van den Heuvel · Accepted solution]

SEP, 9MB/sec is fine for a 100mb setup, but disappointing for a Gigabit setup.

Without actually trying to measure performance, we frequently get 55+ MB/sec with a basic NFS serving setup, makig no special provisions for input data on source this or anythign else.

In a testbed I currently work with we have an archive of 30+GB compressed backups on one
system (where the data is created/massaged) which we uncompress + restore to the 300+GB database files in well less than an hour, this over a gigabit local connection.

Now on Tru64 we could switch our Gigabit links (mostly altion card/driver) to 'jumboframes' to step up from the usual 1500-ish byte ethernet packages to 4000+ bytes. Is that available on hpux? I never checked that out.

Hein.

[Dave Olker]

Hi all,

I've done some performance testing with a pair of Itanium systems running 11.23 and I've been able to achieve near wire-speed Gigabit throughput over NFS. (Repeat, over NFS - not just network traffic, but NFS throughput at near GB wire speed).

For the skeptical out there, here is a sample iozone output of a single process writing 10GB of data across an NFS/TCP connection using a Gigabit connection:

Run began: Mon Aug 9 10:52:06 2004

File size set to 10485760 KB
Record Size 32 KB
Include fsync in write timing
Include close in write timing
Command line used: iozone -s 10g -r 32k -e -c -i 0 -t 1 -f /hp-1/file
Output is in Kbytes/sec
Time Resolution = 0.000001 seconds.
Processor cache size set to 1024 Kbytes.
Processor cache line size set to 32 bytes.
File stride size set to 17 * record size.
Throughput test with 1 process
Each process writes a 10485760 Kbyte file in 32 Kbyte records

Children see throughput for 1 initial writers = 104688.07 KB/sec
Parent sees throughput for 1 initial writers = 104685.68 KB/sec
Min throughput per process = 104688.07 KB/sec
Max throughput per process = 104688.07 KB/sec
Avg throughput per process = 104688.07 KB/sec
Min xfer = 10485760.00 KB

That's over 104 MB/sec NFS throughput.

You'll notice I'm using the -e and -c options of iozone so I'm not cheating by getting the results of storing the data in buffer cache. Also, these systems only have 4GB of memory in them and buffer cache is sized at 15% (dbc_max_pct), so a 10GB write all but nullifies the effects of the buffer cache.

My client is a 2-way rx2600 system and my server is a 4-way rx5670 system. I'm using a VA disk array on the NFS server to hold the filesystem.

Again, this is a single writing process on the rx2600 sending 10GB of data across an NFS/TCP connection to the rx5670 via a single GBE connection separated by an HP ProCurve 2848 switch.

Of course, your mileage may vary...

Regards,

Dave

[Dave Olker]

Hello Hein,

Yes, HP GBE interfaces support jumbo frames. Of course,the switch your connecting to has to be configured to support jumbo frames as well.

That reminds me, the NFS throughput numbers I posted above were not using jumbo frames - standard 1500 MTU size.

Regards,

Dave

[rick jones]

NIC speeds are expressed in units that are powers of ten. Byte sizes (ignoring disc manufacturers :) are generally units that are powers of two.

You can divide 1 Gbit/s by 8 bits per byte to arrive at a maximum of 125 million bytes per second. A MB as a power of two is 1024^2. I _think_ that to go from millions of bytes to MB one would multiply by ~.9537which yields ~119 GB/s.

Now, with a 1500 byte MTU, and ass-u-me-ing you are using TCP and doing say FTP, that is 1460 bytes of data in every 1514 bytes of data on the wire, which then takes "speed of light" as it were to ~114 MB/s, or ~0.1123 GB/s.

Assuming, of course, I've not slipped some digits somewhere.