Ghost Solution Suite

 View Only
  • 1.  Network Infrastructure/Performance Issues

    Posted May 08, 2006 07:52 AM
    We are planning an upgrade to our network and replacement of servers. I want to use all CAT 6 or 6E copper cable/connectors/nic's throughout and I like the features of 3com's 3870 SuperStack series of switches. The plan is to use 5 24 port 3870 switches and be able to expand later. Questions I have are: any suggestions as to shielded versus unshielded connectors/cable and if there is a better switch to use. I am looking at either a 2 or 4 port link aggregating type of TNIC network card (alacritech.com). for better network performance, anyone have experience with this type of network card? Will using multiple or dual core processors have any performance benefit for any Ghost or unicasting functions? Since I use ghost unicasting almost daily in either of my 5 training labs the main goal is the best performance/reliability for Ghost unicasting image transfer speed over a network. Secondly, since funds for infrastructure upgrades most probably will not be made available for at least 8-10 years, I want to install the most up to date/expandable network hardware for at least halfway decent performance in the future. Thanks, Harry


  • 2.  RE: Network Infrastructure/Performance Issues

    Posted May 08, 2006 05:33 PM
    are there any reason why you are not using multicasting?
    multicasting is better when you deploy more than 1 machine at the same time.
    Also ghost multicast is one of the finest multicast tool available in the market.
    So I think, if you need the best performance you could try multicasting.


  • 3.  RE: Network Infrastructure/Performance Issues

    Posted May 08, 2006 08:19 PM
    Unfortunately we can't really recommend specific models of switches or routers to work with Ghost, but the 3Com 3870 certainly does have an impressive feature set. Although it isn't a full Layer 3 switch, the data sheets indicate it does have the necessary IGMPv2 snooping and query generation which should in principle make it a capable multicast performer.

    Probably the main issue I have heard of in dealing with switch and router selection is that some people do have problems when mixing vendors because of slight differences in the implementations of the newer control protocols. As long as the 3Com units stack as widely as you need that shouldn't be an issue.

    > Will using multiple or dual core processors have any performance benefit for any Ghost or unicasting functions?

    Although it's conceivable that they could in future help with CPU-intensive data compression operations during image capture, at present Ghost isn't written to take particular advantage of such features. The image compression is performed only at the client end at present, not at the server. The corresponding decompression operations are performed at the client end when distributing compressed images, and they are not very CPU-intensive so I would expect relatively little benefit to be found in that case.

    In most circumstances, performance of image distribution operations in 100Mbps networks is constrained by the network bandwidth, while in 1Gbps networks the limiting factor is generally the write throughput of the disks in the client machines (which is, with SATA 2 disks in particular, generally rather lower than the headline data rate of the SATA interconnect).

    > I am looking at either a 2 or 4 port link aggregating type of TNIC network card (alacritech.com). for better network performance, anyone have experience with this type of network card?

    Not particularly with Ghost at the server end; but in general, once you move from 100Mbps to 1Gb networks, most image distribution operations seem to be primarily controlled by the performance of the disk subsystems at the client ends. CPU load on the server side during GhostCast is something we do look at from time to time, but we haven't heard any reports of it being a specific customer issue.

    That said, if you are considering storing the Ghost images for distribution on some form of network storage (for instance, an iSCSI volume) then there may well be an argument for a TOE-type card for the path to the network storage on the server machine. Given that the iSCSI market in particular is rapidly evolving it's hard to make definitive statements about performance issues at all.

    The same can be said for the RAID market (especially SATA raid) as well; quite what the "best" storage architecture for the image store on your primary image server is not only hard to say now, but likely to change rapidly as the disks, the controllers and the servers all evolve. The SATA RAID environment is particularly exciting.

    Really, the single most important thing to look at is having a 1Gbps path from the server out to all the clients; this is the part you'll have to live with for the longest, but if you're currently running a 100Mbps network it's going to be a substantial improvement. The choice of e.g. network card or RAID engine for the server machine is not going to make the same level of difference to your environment, and until you have the 1Gbps infrastructure in place you won't really be able to tell where the performance bottlenecks on the server side are, and whether your should concentrate your money more on your server storage subsystem. You do have a lot of exciting options for storage, whether it's directly-attached SATA RAID on the server (which is becoming very affordable) or network storage.


  • 4.  RE: Network Infrastructure/Performance Issues

    Posted May 09, 2006 07:58 AM
    Each lab client is unique and multicasting is not compatible with the scripted/canned oem builds that we use. It's not the Ghost software, it's the oem builds, they are always a "work in progress." Classroom clients can be setup as domain controllers, backup domains, W2K Pro, XP pro, and W2K server in the same training class. We have tried multicasting, and halfway through a 6 week training period the builds deteriorate. We were able to use multicast in groups of 3 as each group is similar but it takes too much time to create/troubleshoot 5 groups of 3 for 20+ classes in each of 5 labs, all while classes are in progress. The unique image per client, unicasted, has proven to be the most reliable for us. I am planning to learn/implement the Ghost Console which should save time. On small builds, (Windows and a small app), imaging 15 clients by using a bootable CD has proven to be faster than retrieving it from a server over an antiquated 10mb network. I am using removable drives so a new build can be worked on or imaged in between classes. Space is limited so there is not a work lab or testing area, everything is built/imaged in the lab.