Preliminary Summary of the test results, Hub vs. Switch

[ScottMac]

Alrighty. The following are the Chariot results, summarized to the bare minimum (to conserve Anandtech's space), of the Hub vs Switch testing from this afternoon. There might be a few surprises, maybe not...

Edit: Complete test results

Edit#2: Jack's hub arrived, we'll do the rest of the tests in the next day or two (6/19-6/20). Jack's hub is a four port, we're gonna test it, then cascade it to the other four port hub & try that too.

JackMDS' 8 port hub has not arrived, when it shows up, we'll expand the test as soon as possible. We were limited to a four port 10/100 hub, and one port had to be the Chariot console (total three machines under test).

The "Failed" tag means the test failed to complete properly due to hardware limitations of the device under test (which happened to be a switch, the hub hasn't done this test yet, but will be tested when possible).

*********
All values are AVERAGE throughput, in Megabits per second. All hub tests were done at 100 Mbps, Half Duplex. All switch tests were done at 100 Mbps FULL DUPLEX (we were running short on time).

The last number in the column is the total throughput (AVERAGE).


*********

Workstation to Server, hub, 1 stream, TCP :89.520

Workstation to Server, hub, 1 stream, UDP :35.301

Workstation to Server, switch, one stream, TCP :94.134

Workstation to Server, switch, one stream, UDP :60.780

Workstation to server, hub, two streams, TCP

29.677
59.757
-------
86.317

Workstation to server, hub, two streams, UDP

36.796
45.417
-------
81.458

Server to Workstation, Hub, two streams, TCP

44.761
44.835
-------
89.492

Server to Workstation, Hub, two streams, UDP
Average (Mbps)

45.250
45.531
-------
89.701

Workstation to server, switch, two streams, TCP

47.520
47.432
-------
94.835

Workstation to server, switch, two streams, UDP

36.429
36.554
------
72.838

Server to workstation, switch, two streams, TCP

47.443
47.491
------
94.851

Server to workstation, switch, two streams, TCP

46.789
47.792
------
93.547

Server to workstastion, Switch, two streams to the server plus one stream WS to WS, TCP

47.482
47.444
92.534
------
186.302

Server to workstastion, Switch, two streams to the server plus one stream WS to WS, UDP

47.412
47.550
57.873
------
152.015

Two Workstations to server plus one stream workstation to workstation, switch, TCP

47.497
47.649
43.836
--------
138.555

Two Workstations to server plus one stream workstation to workstation, switch, UDP

36.405
36.543
57.700
-------
130.116

Workstation to server, switch, three streams, TCP

(Failed)

Workstation to Server, switch, three streams, UDP

31.102
31.099
29.668
-------
91.666

Server to Workstation, switch, three streams, TCP

(Failed)

Server to workstation, switch, three streams, UDP

31.709
32.157
31.782
-------
93.769

******* End (for now) ******

SmartBits: (Summary)

With four streams (four in, four out)the eight port switch transmitted 100% of the generated traffic, at or near "wire speed," with latency figures comparable to many/most/all "commercial" switches.

Throughput was not tested on the hub. The Smartbits "doesn't do" throughput testing on hubs (results show up as 140%+ throughput). Hubs were tested for latency and packet drop only.

SO, there ya go ........


Scott

Thanks to Doug, Eric and Dave for helping.

 

[Zoltarc]

How interesting. The hubs were slighty faster (in most cases)

 

[nexus9]

Not sure what you're trying to test/prove here... the switch and the hub should be similar if you're sending/receiving to 1 machine. Now if you did something like:

Server A <--> Client B
Server C <--> Client D
Server E <--> Client F

All doing transfers of (for example) large files at the same time, the switch should beat the hell out of the hub for total bandwidth. If all of your clients are just going to one machine it kind of defeats the purpose of having a switch...

-Nexus9

 

[ScottMac]

I agree, in a real user environment, all computers talking to one destination (all ports at the same speed)is a bad place for a switch (rehash, old thread)...in the testing environment, it's meant to stress the device under test, mostly for buffering capability, to see how long a buffered packet is held in-queue, and how well the processor and logic respond to the stress.

The last couple tests on the current list have one or two hard streams to a single resource, with another stream from WS to WS. If the processor or memory are not up to snuff, the WS-WS stream would choke (didn't happen until three hard streams).

There's more tests to do, if you want to see something, let me know and we'll try to work it in.

FWIW

Scott

 

[spidey07]

<< How interesting. The hubs were slighty faster (in most cases) >>



huh? i'm seeing the switch as being faster in most tests?

 

[ScottMac]

Bump

^^

 

[spidey07]

SCOTT,

great job!

Is there somekind of conclusion coming as well?

 

[ScottMac]

Conclusions after all the testing is done, but I think the surprises are over.

Didja make it to the web site? Whaddya think about 80-90Mbps on the hub (Mister
&quot;Get a switch&quot; man.....)?

I still like switches, but I don't think hubs are quite dead yet...

Hopefully we can finish tomorrow or Wednesday, then we'll get the whole banana posted and be done with it.

Thanks for looking, responses have been scarce.

Take care

Scott

 

[spidey07]

year is 2001

i'll wait for the conclusions before I EVER admit I'm wrong.

 

[Woodie]

So. When you gonna test my t/r network, hunh??


j/k.

No point in testing, I have no reason to put out any $$ for that enet &quot;stuff&quot;.

Nice job. I'll keep an eye out for the full results, and the write-up.

--Woodie

 

[ScottMac]

Um, actually we did some testing on T/R and T/R over ATM a couple years ago. We even tested TR ELAN-to-EN ELAN through the IBM/Xylan MSS (old Proteon box). I still wake up in a cold sweat screaming from that one...thousnds of nodes, one ATM backbone, doing translational bridging in the ELAN....

Anyone out there doing ATM? How about multiple ELANs through a router? I have some BAAD news....

FWIW


Scott

 

[spidey07]

yeah, put in some ELAN stuff when gigE wasn't available. Even argued the merits of ATM vs. Ethernet. I saw the light as soon as L3 switching came about.

As far as I'm concerned ELANs are the work of the devil. He came up here and said &quot;thou shalt make a LAN technology based on a non-broadcast layer2...thou shalt call it ATM. thou shalt make many different complicated boxes and call them things like BUS, LEC and LES. thou shalt have many sleepless nights figuring out the simplist problems&quot;

damn devil.

BTW - judging from your data a switch is still the top performer???



<< TR ELAN-to-EN ELAN through the IBM/Xylan MSS (old Proteon box). >>

yuck

 

[ScottMac]

ATM ain't that bad, I like it...actually, I'm probably one of the last ATM bigots in the building. I have ATM at home, I use ATM at home, It's wunnerful. I figure if you're gonna use a switch, use a REAL switch....besides, in a year or two when the rest of MPLS gets ratified, it'll be where ATM was YEARS ago..and MPLS will STILL work better over ATM then Ethernet for a couple years after it starts to hit the street (not the current MPLS...the end-to-end stuff).

Most of the anti-ATM guys tend to be Cisco-centric (non-WAN/Strat) types...Cisco HATES ATM...the stuff they put out in their classes in some/many/most cases, is dead wrong (at least for non-Cisco ATM).

I won't say anymore about it (without provocation)...I don't want another Hub/Switch thing to pop up.

Regarding the switch we used in the test, we were all surprised at how well it performed under standard loads (zero packet drop, wire speed (more or less), and latency as good as or better than many/most commercial-grade boxes). Under max loads, it totally flaked at three streams...so (we'll check this out this week)...the hub will probably get more data through, especially in a many-to-one, than the switch (three hard streams or more).

We'll see, the fat lady is still tuning up ....

Take care

Scott

 

[EasyE]

Scott,

Good write up... man my little 4 port hub rocks.

Been busy tweaking my new box.

See ya... E

 

[ScottMac]

Just a quick update: Things got busy at work today, we're gonna do what we can to finish the testing tomorrow (Wednesday) and get the final results &amp; stuff posted by the end of the week. Sorry for the delay. I have some of the Smartbits output, I'll be adding that to the web posting starting tonight.

Thanks for your patience....

Scott

 

[ScottMac]

I put up some SmartBits results on the web page. A couple of the files were corrupt, so I could only post hub latency (0.6us), and switch packet drop (zero across-the-board).

I'll get some more up tomorrow night. We're probably going to finish the tests tomorrow (Wednesday 6/20).


Scott

 

[spidey07]

SO!!! no packet loss on the switch? surprising indeed. i'd figure you could easily overdrive a egress buffer in there.

maybe the SOHO switches don't suck as much as I thought.

 

[L3Guy]

The switch did very well. I did not expect the numbers we got.
These are very fast for any switch. I hesitate so say this is typical of small switches,
but this one ROCKS!

By Memory: (always a iffy proposition)
Smartbits result for switch:
Cut through Latency 7.5 uS at 100% utilization, 64 byte packet.

SF Latency 2.5 uS at 100% utilization, 64 byte packet.
SF latency did not change with packet size on the 4 tested sizes: 64, 512, 1024, 1518 all 2.f uS

4 ports 100Mbps to 4 ports 100Mbps - 0 packets lost.

Doug

 

[spidey07]

impressive....must be asic driven...or some other intelligence in hard circuits and not soft.

rember, netgear was spun off from Bay Networks as their small business/soho line so they've got some lineage.