Samba performance (small correction)
Roeland M.J. Meyer
rmeyer at mhsc.com
Wed Oct 1 18:42:32 GMT 1997
In a practical LAN you lose a lot of packets to collisions. As maximum
traffic loads are approached the collision rate goes way up. Someone, I
forget who and my library is being 're-aranged' at behest of my wife
(direct orders, actually), actually did a statistical proof that this
limit is just a hair above the 50% mark. This proof was presented at the
old 5Mbps standard and caused a lot of argument. With normal routing, the
technology hasn't really changed since then. Only the base frequency was
increased. Only with FDX switching technology can one hope to approach the
baseband throughput, under heavy load. The problem is partially caused by
the CDMA nature of an ethernet. But, these are second-order considerations.
As someone else pointed out to me (picking nits) two machines, on an
isolated LAN, can approach the maximum values. This is exactly what
switching technology does. It makes the individual node believe that it is
talking to exactly one machine, directly. With two FDX capable NICs,
10baseT, and a single cross-over cable, one can achieve up to 98% of
maximum theoretical transfers, on one-way transmissions. However, the one
is not likely to see such configurations outside of SOHO and the lab.
Anywhere else, you need switching technology to see this performance, on a
My mis-statement was in saying that my procedure gave theoretical maximum
results. This is not strictly correct. It would be more accurate to state
that they yielded maximum practical results, if you use the HDX numbers.
The RAW FDX numbers are still maximum theoretical values as one simply can
not exceed the baseband frequency, on a baseband LAN.
I do these calculations all the time as a Systems Architect for
Client/Server systems. I've found that using the HDX numbers gives me a
comfortable engineering margin. Using the FDX numbers always leads to
saturated, and slow, LANs. Commercial interests have a tendency to overload
their systems by as much as 200%, over time. Using the HDX numbers leaves
room for a small bit of growth. The truth is actually somewhere in-between.
But, as I said, this was only a first-order approximation!
At 08:14 PM 9/30/97 +1000, Roeland M.J. Meyer wrote:
>At 06:38 PM 9/30/97 +1000, Bond, Jeffery wrote:
>>On my samba server, which is an ancient 486DX4 with only 8Mb, running
>>FreeBSD, I usually see about 500KB/s when transferring large files over
>>10Mbit ethernet. Only one or two clients are usually connected. I've no
>>idea about NT, because I've never used it, but 500KB/s seems fast to me.
>>What is the max 'raw' throughput of 10Mbit ethernet anyway? About
>To calculate maximum theoretical through-put, do this; Note that this
>procedure only produces a first order (back-of-the-napkin) approximation.
>base = 10 Mhz = 10 Mbps (bits-per-sec)
>FDX = base / 8 = Raw = 1.25 MBps (bytes-per-sec)
>Note: the above is raw bytes per second (Bps), one-way. The only way you'll
>actually see this is with an Full-Duplex(FDX) 10baseT switch. Ethernet is
>normally Half-duplex(HDX) so the practical through-put is half of the Raw.
>Allowing for normal operations (co-ax baseband.[10base2])
>HDX = Raw / 2 = NetSpeed = 625KBps
>This is completely ignoring protocol over-head as well as other traffic on
>the wire. Your 500 KBps indicates;
>HDXoverhead = (625K - 500K) / 625K = 20%
>FDXnet = 1.0MBps
>HDXnet = 500KBps
>This is actually quite good, real close to theoretical maximums. The other
>fellow, talking about 800KBps, must have been talking about a 100baseTX LAN
>with other traffic slowing things down, or they had a FDX switch, rather
>than a HUB/router. Still they had to have other traffic. Basic ethernet
>just can't go that fast.
>A long time ago, when I was doing OSI protocols for Retix, we were
>seriously worried about this through-put, on IBM XT's, being more than the
>processor could handle. We actually slowed the stack down to avoid packet
>retries. The practical limit, for an IBM XT, was 600KBps, bus through-put,
>with 8088 at 4.77 Mhz clock-rate. If we had an ST-506 disk access, at the
>same time, we were definitely going to miss packets, no matter what we did.
>This was around 1983-4.
>However, with 100baseTX, this is a little faster;
>base = 100 Mhz
>FDX = 12.5 MBps
>HDX = 6.25 MBps
>Assuming 20% overhead
>FDXnet = 10MBps
>HDXnet = 5 MBps
>Compared with a T1, assuming the same 20% overhead;
>Note that FDX figures for T1's are irrelevent.
>base = 1.54 Mhz
>HDX = 96.25KBps
>HDXnet = 77KBps
>Note: I have NFS between my two Linux servers. Someday, I'll do some
>through-put comparison testing.
Morgan Hill Software Company, Inc.
Roeland M.J. Meyer, ISOC (RM993)
e-mail: mailto:rmeyer at mhsc.com
Colorado Springs, CO - Livermore, CA - Morgan Hill, CA
More information about the samba