Re: Beowulf Cluster is very slow. Suggestions needed to increase the speed.

Hello again,

Well.. with 10Gbit.. it's an overkill :D with this hardware lab.
Imagine that good 48-port 10Gigabit switch cost around 25-30k US $ (in 
words: thousands). You'll also need 10Gigabit cards (let's say 500 US $ 
each). I didn't saw 8-port 10GbE switch anywhere, if You found it please 
- let me know! :D

Shouldnt You stay at Gigabit? but with good performance switch.
It means high values of:
- switching capacity (in Mpps)
- switching bus (in Gbps)
and low value of:
- switching latency (fractions of ns)

Less expensive way is.. Infiniband network. It usualy starts at 10Gbit 
speeds! There are Infiniband adapters and switches available on Ebay (I 
own two of them) on PCI-Express x4 interface. Some, not all, support 
TCP/IP networking as usual NICs.
This interface is specially designed for clusters due very low latency 
times and low CPU usage at high speeds. But all cluster nodes have to be 
close to each other since cables ARE expensive :)
They also support more equipment than networking (for example: storage).
But on beginning make sure that Infiniband card is well supported under 
Linux, my isn't.

Regards,
TooMeeK




W dniu 2014-10-10 05:34, suresh kannan pisze:
> Hi,
>
> The link you have provided me was very helpful. I haven't prepared
> myself before starting clusters. Few years back i did clustering with
> three system using PVM for different application i.e protein ligand
> docking. With three systems it took some 2 hours to finish a particular
> library screening. Since, it is few hours i was happy and haven't
> checked the speed of the clusters. I thought to myself i learned a
> clustering technique. Therefore, this time i started blindly before
> preparing clustering for different applications which takes months to
> complete a specific job. Therefore, now i learned depending upon the
> purpose of clustering requirement of things will be different.
>
> I learned that speed of the network is a biggest bottleneck for
> clustering especially for our need. In our lab I have found a 1000Mbs
> switch which presently reduces some time compared to 100Mbs however it
> is not efficient. Still it takes a lot of time. I assume we need
> 10gigabit switch. I was not aware of the price these "Gigabit routing
> switch supporting layer 3". It seems to me that even 8-port 10 gigabit
> switch cost approx. 800$. I am still hesitant to ask my mentor for the
> 10 gigabit switch. Since, I dont have no experience i had a thought what
> if i am wrong somewhere although gigabit switch have to work. Now, I am
> looking for 10 gigabit switch in near by labs so that i will connect and
> check whether it sufficiently efficient for our job and then i can ask
> my mentor with some confidence.
>
> i also learned from someone we can use cluster OS specialized designed
> for this clustering purpose for instance Pelican cluster
> http://pareto.uab.es/mcreel/PelicanHPC/. I will install that cluster os
> and check the performance too. I am reading materials to increase the
> speed and performance of cluster in terms of hardware as well as the
> software's. I will tune our application according to our need.
>
> And, also in our lab room we have dual boot systems for 8 people
> (windows with either linux mint, ubuntu, fedora). In given time those
> systems will be either in windows or linux environment depending upon
> our work. Daily 6 hours (nights) and on Sundays our system will be idle.
> I am not capable to use those free computer timing for our advantage due
> to my skill and also time. I guess we can use these computer time if we
> have gigabit switch with networking skills. If someone did similar
> stuffs please document either in your blogs or in email lists so that
> users like me might get benefited.
>
> Although, it took time i learned a lot during this clustering. Thank you
> for your time.
>
>
> regards
> Suresh
>
>
>
>
> On Wed, Oct 8, 2014 at 7:38 AM, Tomcio <[email protected]
> <mailto:[email protected]>> wrote:
>
>     Hello,
>
>     been here since years, but very low number of questions so I'm so
>     happy to see new thread here ;)
>
>     I'm not boewulf cluster admin but..
>
>     I belive Your network is slow because this routing device is
>     internally limited to low throughput, You need Gigabit-capable
>     routing device (read this as: Gigabit routing switch supporting
>     layer 3).
>     I suspect all Your class rooms are in different IP subnets?
>     If yes then the routing device is bottleneck..
>
>     Even modern Core i5 based linux router will create additional delay
>     times in routing when it comes to cluster network, of course under
>     high load at high throughput (what clustering requires). So simple,
>     raw networking is best option.
>
>     Example capable switch models that can do routing are:
>     Cisco SG500X
>     Cisco Catalyst series
>     Dell PowerConnect 7000 series (or lower)
>     Of course, they could be too expensive so just look around on switch
>     that can do IP routing. There are even 8-port Gigabit Managed
>     switches available on market.
>
>     If Your sys admin isn't blocking other network subnets You could set
>     static IP on all Your cluster nodes to different network subnet
>     (let's say.. 192.168.99.1-192.168.99.10 with netmask 255.255.255.0)
>     and check if they see each other (probably this isn't possible) and
>     use one of them to provide Internet access (master node?) with 2
>     Gigabit cards.
>
>     USB Gigabit cards are most up to ~480Mbit in throughput since
>     they're USB 2.0. It's half of needed performance and! be aware of
>     missing Linux drivers.. ;)
>     I see all Your hardware is Gigabit, so I belive You need Layer 3
>     Gigabit switch. Also, please check are Your NICs supporting offload
>     functions, then can help with high network load.
>     I personally these in /etc/rc.local
>     echo "Setting offload functions on Intel PRO/1000 NICs..."
>     ethtool -K eth1 rx on tx on sg on gso on gro on tso on
>     ethtool -K eth2 rx on tx on sg on gso on gro on tso on
>
>     And there are tips:
>     http://cs.boisestate.edu/~__amit/research/beowulf/beowulf-__setup.pdf <http://cs.boisestate.edu/~amit/research/beowulf/beowulf-setup.pdf>
>     See the section: 1.2 Networking Hardware
>
>     Hope this helps.
>
>     Cheers,
>     TooMeeK
>
>
>
>
>     W dniu 2014-10-07 05:50, suresh kannan pisze:
>
>         I am an Indian student in suwon, korea. I built a Beowulf cluster
>         (system information below) with four systems in our lab for our
>         simulation work with the help of good tutorials. In those
>         tutorials they
>         have mentioned all the system should have static ip addresses.
>         Unfortunately, in all our labs we have been provided with dynamic ip
>         address[5 ips for 15 members in three separate labs]. I have
>         requested
>         four more ip's from our university system admin. Due to the language
>         problems, i conveyed the requirement through my korean lab mate
>         and i
>         dont know the reason why he denied us the static ip. So i found
>         another
>         way to skip this procedure
>         http://www.reddit.com/r/__linuxquestions/comments/__2gubad/why_static_ip_address___is_necessary_for_linux/
>         <http://www.reddit.com/r/linuxquestions/comments/2gubad/why_static_ip_address_is_necessary_for_linux/>.
>
>
>         Someone suggested to use a router (one static ip) and set static
>         ip for
>         the four computers through a router. I did that and it worked.
>         However,
>         the cluster is very slow. For instance If i submit my simulation
>         job in
>         a single computer [4 core processor], it takes 2 months to
>         complete a
>         specific job. Although, if i connect 4 systems it shows it take
>         6 months
>         to complete the same job. It is actually using 10 core processor
>         [3,3,2,2-100% each]. I used TOP command to see how much
>         processor the
>         head and other nodes are using. I have used openMPI to do
>         parallel the
>         systems. I am using GROMACS (Parallelization based on MPI has
>         been part
>         of this software). I followed a parallel configuration for the
>         Gromacs
>         with the help of this tutorial
>         http://flakrat.blogspot.kr/__2013/04/how-to-compile-__gromacs-461-with-openmpi.html
>         <http://flakrat.blogspot.kr/2013/04/how-to-compile-gromacs-461-with-openmpi.html>.
>         After reading few posts
>         http://www.reddit.com/r/__linuxquestions/comments/__2gbgbg/what_would_be_the_best___linux_distro_for_folding/
>         <http://www.reddit.com/r/linuxquestions/comments/2gbgbg/what_would_be_the_best_linux_distro_for_folding/>
>         i suspected the network router might be an issue.
>
>         Can you suggest me how can i troubleshoot this problem? Some one
>         suggested to use 2 network ports and make linux as a router and
>         use a
>         gigabitswitch to get the speed. However, we dont have 2 network
>         ports
>         system. If this is compulsory i can buy network ports (USB one).
>
>         Where do i start now?
>
>         Can i make my head node as a router and use USB network port
>         (for the
>         second network port) and connect to a gigabitswitch (any model
>         suggestion?) to connect other nodes. I dont know much about
>         networking
>         stuffs. It will be helpful if any experts can suggest to
>         troubleshoot
>         this issue.
>
>         Thank you for your time.
>
>         regards
>
>         Suresh
>
>
>         System Informations
>
>         Head node Processor : Intel core i3
>         RAM : 1 GB
>         No. of processor : 4
>         Network cards : 03:00.0 Ethernet controller: Realtek
>         Semiconductor Co.,
>         Ltd. RTL8111/8168/8411 PCI Express Gigabit Ethernet Controller
>         (rev 06)
>         System company : Samsung
>         Architecture : x86_64
>         OS flavour : Linux Mint 17 Qiana
>
>         Node1 Processor : Intel Quad core
>         RAM : 3 GB
>         No. of processor : 4
>         Network cards : 03:00.0 Ethernet controller: Realtek
>         Semiconductor Co.,
>         Ltd. RTL8111/8168/8411 PCI Express Gigabit Ethernet Controller
>         (rev 02)
>         System company : TG DREAMSYS
>         Architecture : x86_64
>         OS flavour : Linux Mint 17 Qiana
>
>         Node2 Processor : Intel core i3
>         RAM : 1 GB
>         No. of processor : 4
>         Network cards : 03:00.0 Ethernet controller: Realtek
>         Semiconductor Co.,
>         Ltd. RTL8111/8168/8411 PCI Express Gigabit Ethernet Controller
>         (rev 06)
>         System company : Samsung Architecture : x86_64
>         OS flavour : Linux Mint 17 Qiana
>
>         Node3 Processor : Intel core i3
>         RAM : 1 GB
>         No. of processor : 2 Network cards : 02:00.0 Ethernet controller :
>         Qualcomm Atheros Attansic L2 Fast Ethernet (rev a0)
>         System company : JOOYONTECH
>         Architecture : x86_64
>         OS flavour : Linux Mint 17 Qiana
>
>         Router Company : ipTIME N604R
>         Maximum speed : 160Mbps (LAN to WAN)
>
>
>
>     --
>     To UNSUBSCRIBE, email to debian-beowulf-REQUEST@lists.__debian.org
>     <mailto:[email protected]>
>     with a subject of "unsubscribe". Trouble? Contact
>     [email protected] <mailto:[email protected]>
>     Archive: https://lists.debian.org/[email protected]
>     <https://lists.debian.org/[email protected]>
>
>