Slashdot Mirror


Supercomputers - Does the Cabling Matter?

papaia asks: "Having watched, for a while, the development in the area of high-density server hardware solutions (i.e. blade servers), like IBM's 'top gun', and their increased presence in Data Centers, I have been wondering if anybody has had any experience (thus comments) in regards to how important - in such highly priced solutions - is (or could be) the [always neglected] cabling, connecting the servers. One such comment caught my attention, in this regard. Slashdot, how important is the server cabling infrastructure in your Data Centers, and how do you resolve the cable management aspect of it?"

30 of 95 comments (clear)

  1. sentence (flow) by trs9000 · · Score: 2, Funny

    any experience (thus comments) in regards to how important - in such highly priced solutions - is (or could be) the [always neglected] cabling

    im sure you could fit more parentheses and brackets in there! .....you werent even really trying

    1. Re:sentence (flow) by PerlDudeXL · · Score: 2, Funny

      probably some lisp/scheme side-effects.

    2. Re:sentence (flow) by Bootle · · Score: 2, Funny

      Any sentence where you need to break out Order of Operations, right? God, the Internet has destroyed our communication skills. Kinda ironical....

  2. Always neglected??? Speak for yourself... by Anonymous Coward · · Score: 5, Informative

    When you're running a large linux cluster, it is absolutely vital to plan your cabling and get it right. It just is, for access, fault troubleshooting, avoiding restricting airflow. Get cable tray, lay it to each rack site (you have planned out banks of rack sites and planned the airflow to your air con, right? - and please, tell me you have air con???).

    Allow for larger racks so there's room for cables and PDUs. Really.

    Keep things neat with a raised floor, cable tray underneath. A second tray overhead can be handy, but can interfere with moving equipment, so don't fit it unless you'll need it.

    Be aware high speed cables, particularly fibre, have a minimum bend radius - go beyond it, and you have fucked up the cable. Also, be aware of sag due to weight of cables - compress the cable too much with the weight of the bundle against cable ties, and you can damage it.

    Crosstalk is seldom an issue these days, but be wary of laying power and network too close - even if there is negligible interference, you're safer if they're separated by a decent amount.

    Patch panels are useful - use them. Run cables to patch panels and patch panels ->switches, don't go machine -> switch directly (unless you're doing really high-end stuff (for 1GigE copper or fibre, patch panels are fine).

    1. Re:Always neglected??? Speak for yourself... by Bri3D · · Score: 2, Informative

      Watch your raised floors, and use trays though...It can get *really* nasty down there without management and labeling. And yes, use patch panels, patch panels, patch panels. This way you can tell exactly where each port goes and switch switches(no pun intended) easily.

    2. Re:Always neglected??? Speak for yourself... by cymen · · Score: 2, Interesting

      Power down below (and water detection). Everything else up above. High quality labeler.

    3. Re:Always neglected??? Speak for yourself... by photon317 · · Score: 2, Informative


      Another tip - where copper signal cabling (ethernet, serial, scsi, whatever it may be) has to come near power cabling, always cross them at right angles instead of running them parallel to each other, this greatly reduces the chances of inductive interference.

      --
      11*43+456^2
  3. Reliability by Leroy_Brown242 · · Score: 2, Insightful

    I think that it'd be fairly important. Even if only for reliability. Because speed and reliability are equally important.

  4. Cray by Tersevs · · Score: 2, Interesting

    This isnt really an answer - more a bit of trivia:
    The ol' Cray-X supercomputer where round (their cabinets where placed in a circle) so that the length of the cabling could be kept down. Back then the synchronisation between pulses in different cables where a problem. And there really where a *snakepit* of cables between the cabinets.

  5. literally speaking, no by St.+Arbirix · · Score: 4, Insightful

    Literally speaking the cabling won't matter at all. Whether the resistance of your wires is high or low the electrons are going to travel through it at the same rate. What may be a worry is freak occurances of inductance between wires which could possibly mess with your data, but I'm not sure how common that is. The signal is digital so it's going to be either a 1 or a 0 depending on the voltage of the line, and it's usually difficult to make the voltage do something as drastic as go from +5 to 0 or +5 to -5.

    I always have to laugh anytime I go to the store and look at the number of things that say digital nowadays. Wal-Mart sells digital telephone cable for your computer to connect to the wall as if 5 feet of high quality cabling with gold plated tips is going to make a difference in the 30-odd mile trip to your ISP. A friend of mine boasted his "digital" headphones to me once and I had to beat him down to the opinion that "there must be something better about them, so I'm happy" (this sort of pacifist optimism is the bane of IT in general).

    Best Buy sells Monster digital audio cables at something like $20 for a 5-ft cable. I had to argue with my father trying to convince him that the cheap RCA cables we already had back at home would be perfectly capable of communicated a digital signal the 5 inches between the DVD player and the receiver. I could have ripped two wires from a speaker cable to connect the two devices and would have gotten just as good sound.

    People don't seem to want to realize that digital implies lossless or error-corrected. They don't understand that the "premium quality sound" transmitted between devices can be done using the cheapest electronics equipment available.

    I'd save a fortune on car stereo...

    --
    Direct away from face when opening.
    1. Re:literally speaking, no by PurpleFloyd · · Score: 2, Informative
      While the information in your post holds true in the audio world, things start getting a little different when networking comes into the picture. While Monster Cable is quite possibly the single greatest scam perpetrated on consumers of the last 10 years, things change when you move from nice, low-frequency audio into the world of computing.

      As you probably know, standard twisted pair cable comes in several grades, or categories. Cat 3 is the minimum acceptable for 10baseT, Cat 5 for 100baseT, and Cat 5e for 1000baseT. Cables are sorted into categories depending on the highest signal frequency they can pass reliably - 16 MHz for Cat 3, for example. Unfortunately, this means if you were to wire a cluster with Cat 3 wiring and try to run a 100baseT network (at 100 MHz), you would probably get errors in the datastream, resulting in lost frames; this can cause subtle errors that will spawn many headaches down the road.

      The bottom line is, don't skimp on your cables. While Ethernet can be quite resilient (I've seen 10 meg go over a barbed wire fence, albiet slowly and with about 80% dropped packets), skimping on cable to save a few bucks can really cause serious problems down the road. The money you "saved" by installing cables that aren't up to spec will be obliterated tenfold when you factor in the cost of ripping out all the old cable and running new stuff that's up to spec.

      --

      That's it. I'm no longer part of Team Sanity.
    2. Re:literally speaking, no by alienw · · Score: 2, Interesting

      Please, if you don't know what you are talking about, shut up. You are spreading misinformation.

      First, there is no such thing as a digital signal. You can't send numbers through a wire, you can only send voltage levels. This is an example of an analog signal. Poor quality cabling or interference can and will cause errors in transmission.

      Second, nobody uses 0/+5V signaling for anything modern. This is not compatible with high bitrates. For example, USB 2 high speed uses 400mV differential signaling at 480Mbps. Cable and connector quality is critical, and poor quality cable will not work at all.

      Third, most digital interfaces have no error correction capability. Digital audio (SPDIF, which is what you were talking about) has no error correction OR detection capability. If you have bad cable, it will cause sound glitches, crackling, and other nastiness. Also, SPDIF transmits the master clock over that cable. If the cable is of poor quality, it will cause excessive clock jitter, which reduces sound quality and causes distortion.

      By the way, you can't use audio cable for SPDIF. SPDIF requires coaxial video cable (75 ohm impedance). It will not work well with anything else.

    3. Re:literally speaking, no by St.+Arbirix · · Score: 2, Funny

      I've seen 10 meg go over a barbed wire fence...

      So that's how you're supposed to handle physical security for your workstation. My new hero.

      --
      Direct away from face when opening.
    4. Re:literally speaking, no by alienw · · Score: 3, Insightful

      Well, actually, SPDIF requires 6MHz bandwidth. As far as the 75 ohm impedance: I didn't pull it out of my ass, it's specified by the relevant standard (IEC958), and it's not some kind of audiophile debate. Basically, it's the impedance of standard TV coax. Pick up any book about radio and read about impedances. Here's a link to a short description of SPDIF, in case you are curious: here.

      As for your link: don't believe everything you read. The Audioholics article shows many gross misunderstandings. For instance, the reason direction is marked on some high-end cables is for optimal grounding, not because cable manufacturers don't know audio is AC. Also, they seem to have failed physics when they claim that a battery cannot do anything if the circuit is not complete. Ever hear of FETs?

      Cable quality certainly makes a difference for just about any application, including audio. Even a "digital" protocol like USB imposes a number of requirements on the cable quality. Cable quality for sensitive analog signals is even more critical. If you have a few hundred thousand dollars' worth of test equipment, you can probably quantify the differences, calculate bit error rates, and so on. It's easier to just listen, though.

    5. Re:literally speaking, no by themuffinking · · Score: 3, Funny

      No, yourself.

      It matters how many wires are in the cable. A cable is a mass of wires, and the more wires, the more data can be sent at once. Hence the sucktasticness of phone lines as internet connections with their few measly threads of single wires. Now, a broadband cable line has a whole mess of wires in it. A T1 or T3 line has an unfathomable number of seperate cables twisted upon each other within it. It's not the wire that matters, it's the number of them.

    6. Re:literally speaking, no by Technician · · Score: 2, Informative

      It's easier to just listen, though.


      I've seen many people fooled by listening.

      Take two stereo's. One has 0.005% THD and the other has 0.1% THD. Do a side by side test, but have the 0.1% reciever set 3 DB louder. Guess which sounds better?

      The Audioholics article shows many gross misunderstandings

      What I don't understand people who spend $10/foot for speaker cables, then don't want to damage an expensive cable so they have a 30 foot cable to go 5 feet to the speaker.. If you are measuring results of cable inductance, A cheap 5 foot cable of the same conductor size as an expensive cable, is better. Cable length is important. This is more important the more the load impedance does not match the transmission line impedance. A 75 ohm SPDIF cable feeding a 75 ohm SPDIF receiver is a good match and will have minimum distortion and thus few data errors. A 120 ohm impedance $10/foot speaker cable feeding an 8 ohm impedance speaker is going to see high frequency attenuation due to the inductive reactance presented by the speaker cable. So many cable manufactures try to go with low capacitance cables to reduce the shunt capacitance. Low capacitance means a higher impedance because the inductance per foot remains the same. This makes a larger impedance mismatch between the cable and speaker. A high capacitance cable will have the capacitive reactance canceled by the inductive reactance resulting in a lower impedance cable. This would be a better match to the speaker. Any good radio tech knows the best response with the least reflected power is when the load impedance matches the transmission line impedance. The insulation used is important. Low dilectric loss is important to reduce high frequency attenuation.

      Just for grins, visit your local wire and cable outlet and try to find sweep tested 8 ohm impedance wire...

      Where it counts to have quality known impedance cable, it comes sweep tested. They don't sweep test speaker cable because the cable impedance is such a bad match to the load. In this case, shorter is better.

      That's why in the days of RG58 ethernet, the cable was terminated with 50 ohm terminations. The termination did not overload the network cable, but eliminated unwanted reflections that would contaminate the data.

      This is also why pro audio (stage) uses low impedance microphones. The long haul from the mike to the mixer is done using a low impedance microphone feeding a shielded twisted pair cable feeding a low impedance mixer. Typical impedances are 100-250 ohm. Plugging in a high impedance microphone will overload it resulting in poor sound. Plugging in a low impedance microphone on a long cable into a high impedance mixer also results in poor sound unless a matching transformer is used. When matching the cable impedance to the source and load, you get the most power transmission with the least distortion.

      When you go very short distances such as the one you mentioned, (SPDIF) then neither the capacitive or inductive components of a cable are very significant, and as mentioned, a bad match may work OK for a short run. In a short run, cable capacitance and inductance impacts on the signal are small.

      In theatrical lighting, the DMX512 standard specifies the need to terminate the data cable into it's impedance which is near 120 ohm. Many manufactures mention the termination is not needed for short runs. This is in violation of the spec, but they have found it still funtions for runs of less than about 30 feet. Longer runs must be terminated because data errors will cause problems. DMX512 operates at 250K Baud. The 6Mhz SPDIF signal runs at 24X that rate. Cable problems will show up at shorter distances on the higher frequency applications.

      --
      The truth shall set you free!
    7. Re:literally speaking, no by forged · · Score: 4, Informative
      I've seen 10 meg go over a barbed wire fence...

      Happy you got modded funny, however Long Reach Ethernet (LRE) does exactly what it says with very good throughput (we're nowhere near the alledged 80% packets loss of the parent post).

      Oh, and the video clip which shows Ethernet over barbed wire is at the same url on the right-hand side where it says "Video: Charlie Giancarlo Demonstrates LRE Technology". It's nice to see it once for the "Wow!" effect. You'll also see the demo go over Cat3, Cat5, speaker cable, coax and lamp cord...

  6. Crosstalk by macz · · Score: 2, Interesting
    Short of a faraday cage (around each cable) there isn't any way to prevent ALL cross talk, but it is surprising how important using quality cables (Cat 5e or better yet Cat 6) is to reducing overall network latency.

    If nothing else, in an extremely complex environment, if you use a quality cable and quality connectors (skillfully attached) you can eliminate the bus as "one more thing to check" if you are getting unexplained slow downs. It is a nice way to shorten the troubleshooting to do list when you are up to your eyeballs in alligators and the pager wont stop buzzing.

    --
    ...But I digress. TREMBLE PUNY HUMANS!ONE DAY MY SPECIES WILL DESTROY YOU ALL!
  7. Use high-quality cabling, but don't overspend. by mewyn · · Score: 2, Interesting

    Quality cabling always will make a difference. Not quite a computer situation, but still similar, my friend was recently hooking up a new DVD player up to his new projector with component video inputs. He first just grabbed the first pair of RCA cables that he could find. The projector kept resyncing with the YPrPb inputs. Despite soomeone else's refusal to accept it, I told him to pull out some video monster cables. Once he did that, it eliminated the resyncing.

    Cable quality will affect both digital signals and analog signals alike. A bad quality cable will generate a good share of dropped packets, or corrupted data, causing more resends or less accurate data. Also, take care if crimping your own cables, make sure you untwist wires as little as possible, and break the insulation and sheilding as little as possible.

    With that said, don't be like a crazy audiophile (key word here crazy) and spend thousands of dollars just on cabling (I know an audiophile who spent 500 dollars on a 6" cable, when a $25 monster cable has the exact same specs. He claims to hear a difference, but I call b.s. on him.). Spending more means getting better, but only to a point.

  8. Yes, and documentation by rueger · · Score: 3, Insightful

    In any setting the quality of cabling does matter. Or, more specifically, the quality of the connectors on the cable, and the quality of their installation.

    Anyone who argues otherwise should recall that the first step in troubleshooting is almost invariably to check the cables.

    While I am happy to use zip cord to wire my stereo speakers, I wouldn't trust dollar store cables for anything mission critical.

    More important though is to document your cable runs, or even better tag each cable so that you have some idea where it begins and ends. You may know what goes into and out of that big ball of CAT-5 on the floor, but the guy who follows you will have no idea.

  9. You don't want to raise the ire... by mosel-saar-ruwer · · Score: 3, Interesting

    People don't seem to want to realize that digital implies lossless or error-corrected. They don't understand that the "premium quality sound" transmitted between devices can be done using the cheapest electronics equipment available.

    Digital, maybe, but you don't want to raise the ire of the analog stereophiles: You'll get everything from Stereo cables make a difference to Debunking the Myth of Speaker Cable Resonance, not to mention forests worth of dead tree sacrifices for Speaker Cable Face Offs.

    And please, please, please, please: Don't get them started on Solid State -vs- Vacuum Tube...

    1. Re:You don't want to raise the ire... by alienw · · Score: 2, Informative

      You are perpetrating another myth. There is a very good reason to use thick cable for 12 volts. As you decrease the voltage, you need to transmit more current to get the same amount of power. An amplifier that draws 10 amps from a 120V outlet will have to draw 100 amps if run from 12V. Thus, the cable needs to have 10x the cross-sectional area. As far as gold plating: it prevents corrosion and has low electrical resistance, which is important at those amperages.

    2. Re:You don't want to raise the ire... by Glonoinha · · Score: 2, Informative

      The gold plating isn't so much for corrosion resistance or for the amazing electrical conductive qualities - it is because gold makes the best interface between two electrical conduits. When the two wires or connectors come together - gold does this best. As far as the actual wire goes for moving the electricity around silver is better than gold, once the connection has been made - IIRC satellites use gold connectors, but silver wires (since money is no object, and failure isn't an option.)

      --
      Glonoinha the MebiByte Slayer
  10. "Audiophile" cables by pv2b · · Score: 2, Funny

    $500 for a 6 foot audio cable? Your friend prolly got some cheap low-end swill. :-)

    Here's the real good stuff. I wish I weren't making this shit up.

    This previous /. thread has a few other interesting examples, including the one above.

    Me? I just hook everything up using lamp cord.

  11. Re:No by BladeMelbourne · · Score: 2, Funny
    How can I shut down if I can't click on the Start button?

    Windows (TM) Supercomputer Uptime: 23:59:5BSOD!

  12. Cabling is a critical component by Ropati · · Score: 2, Interesting

    Behold the Rat's Nest,

    If your datacenter is 24/7, doing costly (financial), life critical (healthcare) or corporate production, then cabling ranks right up there with A/C and power. In fact all three of these are more important than apps or server platforms.

    I mean, most signal cabling is now part of a network, (IP, FC, ESCON, Token Ring, etc.). A single cable failure can lead to a network failure which, like an A/C or power failure, affects a good portion of the datacenter.

    I've seen poor cabling take out a datacenter on a couple of occassions. In one case, the engineers had loosely laid fiber cable for their network backbone under the computer floor. The cable draped over metalic power conduit. A year later, the datacenter contracted to have the power upgraded. The electrician pulled out the old conduit taking about ten fiber pairs with it. The company lost a good portion of their IP connectivity for several hours. Cabling is critical.

    Cabling should be well thought out and properly run. The best systems I've witnessed are seperate trays under a computer floor for copper signal, fiber signal and a third for power. Cable runs go down the rows under the backs of the racks. All trays have proper feeds for each rack. All new cabling is quoted, and contracted before installation. Any equipment removal entails cable removal.

    The best cable management system I've ever seen was at a TV station. The chief engineer kept several different cable lists depending on the cable function. Each cable was given a number. Once the cable was run, on his inkjet printer, he printed up cable labels using a Brady label sheet. The label identified the use, local connection, remote connection and number. There were never any problems disconnecting or reconnecting equipment.

    Cables tell the story. If you are ever going to contract a datacenter for rack space, a visual check of the cabling will tell you more about the establishment than any brochure or spec sheet. If the cables are well run, you can bet - the power and A/C are properly spec'ed and redundant, their bandwidth adequate, and their building secure environmentally and physically.

    --
    machinator omnis sine licentia
  13. Under the floor? Reconsider! by Myself · · Score: 2, Informative

    I strongly disagree with putting cabling under the floor. Out of sight, out of mind works fine when you're talking a few cat-5 runs to a cubicle farm, but when connections are your main business, put them up where they won't be neglected.

    The under-floor space can be used for AC power, if you use AC, but that's usually just for convenience outlets. Downflow air handling units that use the floor for air distribution are good too. I've even seen installations where the DC power cabling was run under the floor, and it simplified things greatly. But please, don't put your signal cables down there.

    For one, it's easy to drop a tile into the floor while trying to remove it. One bad suction cup can cause the crushing or cutting of a cable. For another, it's awkward to feed cables through the little cutouts at the bottom of a cabinet. I've seen a lot of dirty or damaged connectors because of this.

    If you're not bolting your cabinets to the floor, it also creates a shear point if the cabinet shifts. Please do bolt down your racks and cabinets, because they can tip.

    Hiding the cabling also encourages poor workmanship. When someone has 20 feet of slack to store, and they throw it in a clump under the floor, it's a nightmare when another cable in the same area has to be pulled out. The initial infraction would've been noticed immediately if it'd been overhead, in plain view.

    A well-designed overhead cable rack system is superior to any floor system. It's cleaner, because there's literally less dust colleting on it. Running cables overhead doesn't involve dragging them through a pile of connection-ruining crud. It's easier to install, because you don't have to contend with tile supports. It's easier to expand, because you can visualize the whole layout easily, and see where the congested areas are.

    Furthermore, overhead rack is a natural companion to fiber trough systems, most of which are intended to be overhead. If you have a mix of fiber and copper, and most of us do, you owe it to yourself to plan a system that accomodates large amounts of both. As equipment density rises, the amount of cabling you'll need to bring to each rack also rises. Plan for that.

    Also, plan for slack runout areas. The cables are never the exact length you need. Running them back and forth in the rack can create all sorts of tangle problems. Having a designated path to run your slack loop down can really make tearouts less dangerous.

    Also, don't underestimate the sheer size of the cabling you're dealing with. I saw one particularly bad example, where a company had laid out their aisle of patch panels very carefully. The bottom of each bay was for panels that went to transport equipment, and the top was panels for other equipment. That way, most cross-connects could be made without leaving the bay. There were cable management rings to accomodate the occasional jumper that had to go between bays. It worked great.

    Then they merged with another company, and the recordkeeping system changed. The new system made port assignments automatically, and it didn't respect the physical layout of what was where. Now the majority of jumpers were long, inter-bay runs. Over time as circuits got moved around, the management rings got filled, overoaded, and eventually stuffed to the point that the mass of wire was essentially solid. You could punch the bundle and it would go "thud". Pulling out a jumper was likely to burn through its neighbors simply due to friction, so they stopped pulling old ones out.

    Eventually they added a dedicated piece of cable rack, and run all interbay jumpers up there. It was ugly, and awkward, but it worked. The initial system was much better, but relied on a level of care and planning that the new owners weren't willing to provide. Consider this: Will your successor's successor curse your name, or laud you for laying out a comfortable, expandable environment?

  14. Cabling Strategies by bill_mcgonigle · · Score: 2, Interesting

    It's not the cables, it's how you cable. Kentucky bred a cabling strategy for their cluster. The Big Mac project at VT was supposed to release a software package that made cluster cabling easier, IIRC, but I can't find it anywhere.

    --
    My God, it's Full of Source!
    OUTSIDE_IP=$(dig +short my.ip @outsideip.net)
  15. colors and numbers by smoon · · Score: 2, Interesting

    One thing we do with 'cat 5' cables is color-code different length cables, so black == 3 feet, green == 7 feet, yellow = 10 feet, orange == 14 feet, red == 25 feet, grayish-white == 50 feet, red with yellow boots == crossover cable. This has been helpful in a number of ways. -- it's unlikely a white cable will be to something else in that cabinet, crossovers are easily identified, longer cables are probably for servers further away from the switch/patch panel/whatever, it provides some color distinction in otherwise monochrome patch panels, it's easy to stock and order cables like this.

    One thing I wish we did is have unique serial numbers on both ends of each (and every!) cable. While it's possible to trace cables using the tried-and-true tug-and-feel method, in reality it sucks and printed documentation is difficult to keep in sync with reality.

    I've also seen cables color-coded for other purposes, but these haven't worked as well e.g.: one color is for network, another for KVM, another for switch uplinks, etc. This works well until you need a KVM cable, but don't have the right length in the right color so substitute "temporarily", blowing the scheme completely since 'temporary' is a synonym for 'permament' in most datacenters. another example: Use every color available randomly in the hope that there are only so many hot-pink cables with a green stripe in your datacenter making it easier to trace things. In reality this last example doesn't scale well and makes patch panels look really untidy.

    As far as what I *think* you were asking, which is whether there is some qualitative difference between cables -- there is. Make sure you get 'certified' cables from a trusted vendor, preferably each one individually tested with the results pasted on a sticker on the (sealed) bag each cable comes in. Also make sure you get 'plenum' cables where necessary to comply with fire codes and just plain common sense. I'd say any permament infrastructure cables (not patch cables) should be plenum whether they are legally required to be or not -- if you have a fire you'd be better off without a few hundred extra pounds of fuel to keep it going. Beyond plenum/pvc and tested cables there isn't much else to stress over -- thank god "Monster" doesn't make patch cables with 24k gold connectors to hoodwink unsuspecting people -- if the cable tests good the rest doesn't matter.

    --
    "But actually trying to use m4 as a general-purpose langage would be deeply perverse" --ESR
  16. Re:WARNING: In wall stereo speaker must be insulat by technos · · Score: 2, Informative

    What the gentleman meant is they need to be jacketed.

    For in wall applications, you can't use plain old insulated copper anymore, you require something with a fire-resistant outer jacket.

    Most companies accomplish it with PVC or PVC and foil.

    --
    .sig: Now legally binding!