Tech Arsenal 1/Tech Arsenal (Arsenal Computer).ISO/tek-07/perform.txt

Text File | 1992-10-06 | 24KB | 438 lines

          TECHNICAL BULLETIN                                        PERFORM
          LANtastic Performance Issues

          The  following  information  is  based  on  in-house  testing  of
          LANtastic.   If you  are not a  registered LANtastic  owner, your
          questions  should be  directed to  our technical  sales staff  at
          [phone].   Any questions or comments on  this bulletin may
          be forwarded by  FAX [phone] or mailed  to: Artisoft Inc.,
          691 East River Road, Tucson, AZ., 85704, ATTN: Technical Support.

                               GENERAL HINTS AND TIPS
                     INCREASING PERFORMANCE OF LANTASTIC NETWORKS


          "What can I do to speed  up my LAN?" A very common  question, and
          one to  which there  is unfortunately, no  cut and  dried answer.
          Network  performance depends  on many things.   Getting  the best
          performance out  of your  particular installation  may require  a
          little experimentation.  This bulletin  presents a  discussion of
          some  of the  major factors  that affect  performance, and  a few
          general suggestions for improvement.

          Choosing your Hardware

          When you're choosing the hardware for your LAN, keep in mind that
          the  things  that  make  a  fast  computer  are  also  the  basic
          ingredients of a fast network.  A fast processor, fast memory and
          a fast disk drive  all help to provide better  network speed. For
          best  performance,   you  should   choose  fast,  well   balanced
          components. The two  biggest factors are CPU speed  and hard disk
          speed.

          It's fairly easy to find fast computers and fast disk drives  out
          there in the marketplace, but for the best results make sure that
          the components you buy  are well matched. If  you install a  high
          performance disk  drive in a  4.77 Mhz  XT, your network  will be
          slow because the computer's CPU won't be able to move data around
          fast enough to keep up with the  drive, let alone the rest of the
          network.  Similarly, if you install an old 20mb MFM drive in your
          386-33,  the fast  processor will  spend  a lot  of time  waiting
          around for the old drive, as will the rest of your network users.

          The transmission speed of the  adapter cards you choose will also
          make  a difference. With reasonably fast machines, upgrading from
          a LANtastic 2mbps adapter to  a new Artisoft Ethernet adapter can
          approximately  double  the  throughput  of  the  network.  Again,
          balance  is vital.  If you  have a  slow processor,  switching to
          Ethernet may not make much difference -- the  2mbps adapter has a
          coprocessor  that handles all  the low level  NetBIOS processing,
          whereas the  Ethernet adapter  relies on  the computer to  handle
          everything.

          Optimizing Software Performance

          Now that we've dealt with  hardware selection; let's see what can
          be done  to improve  the performance  of existing  installations.
          First, let's  tie down what  we mean by  the "speed" of  the LAN.
          There   are  really  two  things  at  issue  here,  the  physical
          transmission speed of the hardware, and the effective throughput,
          or how  fast data  gets from  one place  to another  on the  LAN.
          Assuming for now  that the hardware is fixed  and immutable, here
          are some suggestions  for getting the  best throughput with  your
          particular application.

          Optimizing Network Configuration

          There are  two  major configurations  of  the network  that  will
          greatly benefit overall network performance.

          The first is  simply using LANtastic's ability to  have more than
          one  server on the network.   Splitting up  network tasks so that
          different programs  are run on separate servers will decrease the
          load on  each individual  server, this  will improve  the servers
          response (and on a  non dedicated server improve  the performance
          of the computer for the person sitting at the keyboard).

          The next is setting up  a dedicated server (while LANtastic  is a
          Peer to Peer  network - Artisoft considers a  dedicated server to
          be only a  special case of peer to peer  networking). A dedicated
          server can  be specifically tuned to optimize  the performance of
          the  Network portion  of the  computer  at the  detriment of  the
          Workstation (keyboard) portion of the computer.

          This  is done  by using  the  RUN_BURST command  line switch  (on
          AILANBIO with  versions  of  LANtastic  prior  to  4.1,  on  both
          AILANBIO and SERVER with version 4.1 and greater).

          This switch controls the amount  of time that the program (either
          netbios or  server) devotes  to  processing network  tasks.   The
          higher the value the longer amount of time allocated.  Since both
          REDIR and SERVER  release  control if there are no tasks  waiting
          neither  switch  will conflict  with  the other.    Setting these
          switches  to the  max 254  will  give optimum  performance for  a
          dedicated server.

          In a dedicated server, memory no longer is a major consideration.
          Large  sizes  for  the  servers  caching  and buffers  should  be
          experimented with.  (See server buffering below.)

          Another  command that  is useful  for a  dedicated server  is the
          ALONE command  included with  v4.0 and higher.   When  the server
          program  is used to  access a  resource on  the server,  it first
          swaps out all the DOS stacks and control memory.  It does this so
          a  foreground task  can coexist  with the  server running  in the
          background. On  a dedicated  server Alone  tells the  server that
          there will  NEVER be  a program  running on  this machine  in the
          foreground using  DOS and  so server saves  time by  NOT swapping
          into and  out  of DOS.   This  means that  you cannot  run a  TSR
          (Terminate and Stay  Resident program)  on  this dedicated server
          that will use dos calls.

          General Improvements

          FASTOPEN

          Running  the  DOS  FASTOPEN command  will  improve  network speed
          considerably(except  with MSDOS 5.0 because of a reported problem
          in this version's FASTOPEN).  FASTOPEN stores the location on the
          hard disk of each previously  opened file to make repeated access
          to this file much faster.


          SEEK CACHE

          LANtastic 4.1 contains a new cache built into the server program,
          SEEK CACHE.   SEEK CACHE  caches each users position  within open
          random  access files. Typically  a seek cache  will significantly
          affect performance of random access file operations if:

               1> Files being accessed are relatively large.
               2> The files are kept open and accessed frequently.

          The optimum  size for  the  SEEK  CACHE is  the SIZE  of database
          files/cluster  size (typically  4096)* 32 +  the Number  of files
          *32.


          RESOURCE CACHE

          When a  user on LANtastic  accesses a resource, LANtastic  has to
          check  the  resources  Access  Control List  (ACL)  to  see  what
          permissions  the user has.   This  requires loading  the resource
          information  from the  hard disk  and  checking it.   To  improve
          performance  with LANtastic  NOS  4.1  resource  caching  can  be
          enabled on  a server.  This means that  as each resource (disk or
          printer)  is accessed  it's access  information   is stored  in a
          permanent  cache  in the  server.    There  should be  one  cache
          specified  for  each commonly  used  resource.  If  a server  has
          several  resources, but  only a  few are  normally  accessed, the
          cache number can be set to the  number of commonly used resources
          +1.   The  caches  are flushed  and  used on  a least  frequently
          accessed  algorithm so system  performance will be  only slightly
          affected.


          INDIRECT FILE ACL's

          Indirect  files  will  greatly effect  network  performance, with
          LANtastic  NOS 3.0x  eliminate  the I  from  each disk  resources
          Access Control List.   With NOS 4.xx  the I ACL has  already been
          eliminated. With the  I ACL in place  each time a file  is opened
          the NOS checks to see  if the file is an  indirect file.  If  you
          are not using indirect files  eliminating this ACL will save time
          when a file is opened.

          SERVER AUDITING

          Indirectly,  server auditing  can  also have  a  great effect  on
          network performance.  If auditing  is enabled and the audit files
          are not cleared on a regular basis this file can grow very large.
          This by itself is not a problem; but since auditing requires that
          an entry  be made to this file every  time an audited function is
          used,  the  time  spent seeking  to  the end  of  this  file will
          increase as the file size increases.

          SEND SERVER ID

          You can reduce the load on heavily used networks by disabling the
          Send Server ID  selection in NET_MGR, Server  Startup Parameters.
          With this selection  enabled the server sends out  a message once
          every retry period (55ms x retry_period setting), this message is
          just that the machine is a server  and available.  If your system
          is using batch files this message  is not needed.  If the  server
          name  is known, it can  be logged into even  if the server is NOT
          sending out it's ID.

          DISK CACHING

          Adding  a disk  cache to  the  server also  helps performance  in
          almost  all situations.  Disk caching  programs  work by  storing
          frequently  requested  information  in memory.  This  reduces the
          number  of physical accesses to the  disk. Since accessing memory
          is  almost always faster  than accessing the  disk drive, caching
          can greatly improve the disk performance of your network.

          The LANtastic  disk caching  program;  LANcache, is  specifically
          designed to  work with LANtastic networks. It  caches disk writes
          as well as disk reads and when it actually does  have to write to
          the disk,  it  does so  as  a background  operation so  that  CPU
          operations can continue.

          Choosing the Best Buffer Settings

          Beyond these general changes, getting the best performance out of
          your specific  application  is a  matter of  fine tuning  several
          parameters. LANtastic has four sets of buffers, each of which can
          affect performance in specific situations.

          REDIR Buffers

          It takes the network about the same amount of time to send 1 byte
          of data as it does  to send 256 bytes of data. Every  packet sent
          over the  network takes  time to construct  and time  to process.
          Given this, it makes sense to include as much data as possible in
          each packet -- it keeps the network  overhead impact to a minimum
          and  reduces network traffic. The REDIR program accomplishes this
          by buffering network requests at  each machine. Here's an example
          of REDIR's buffering in action:

               You are creating a file of customer names  on the server and
               adding  data to it. You type in the  names one at a time and
               your  program writes  the names  to the data  file one  at a
               time. Rather  than issue a  network send for each  name, the
               data gets stored  in the local REDIR buffer  until you close
               the file or fill  the buffer. At that point, REDIR  issues a
               single request to send the entire buffer.

          REDIR  buffers read requests similarly. If  you request an amount
          of data smaller than the  REDIR buffer, REDIR requests one buffer
          full  of data  from the  server. Then,  if you  request the  next
          sequential piece  of data,  REDIR can supply  it to  your program
          without even making a network request.


          REDIR buffers are  only used for small (small  means smaller than
          the REDIR buffer  size) sequential reads or writes  to files that
          are NOT opened  in the sharable mode. If you read or write blocks
          of  data  larger than  the  REDIR  buffer,  the network  software
          bypasses it and  moves data directly from the applications buffer
          directly to the NetBIOS. The  ideal REDIR buffer size then, would
          be  about the  average  size  of the  packets  that your  network
          typically sends and receives. You  can set the size of the  REDIR
          buffers with the SIZE= command line switch.

          The number  of REDIR buffers  can also affect the  performance of
          many applications. If  you allocate multiple REDIR  buffers using
          the  BUFFERS= command line switch, the network tries to associate
          each file opened  in the non sharable mode.  Network printers are
          handled just like other files.  Ideally, you should have a buffer
          available for  each open non  sharable network file. If  you have
          more files than available buffers, the network flushes the  least
          recently used buffer and assigns into the new request.

          WordPerfect is a  good example of  a program that is  affected by
          the number  of buffers.  If you have  only one  buffer allocated,
          printing can be very slow because WP has to read the document and
          send data  to the printer with  the same buffer. The  buffer gets
          continually filled and flushed, back and forth, with printer data
          and  document data. In  this situation, allocating  just one more
          buffer  allows WP  to use  one  buffer for  document reading  and
          another for printing.

          Server Buffers and Tasks

          A server buffer is at the other end of a  REDIR buffer. When, for
          example, a workstation makes a read request, the server reads one
          server  buffer full  of  data,  beginning  with  the  information
          requested. If the workstation then asks for the next piece of the
          same information, the  server can get it from  the buffer instead
          of physically reading the hard disk.

          The server  buffer size  can be changed  from the  Server Startup
          Parameters option in  the NET_MGR program.  Be aware that  bigger
          isn't necessarily  better -- as  with REDIR buffers,  the optimal
          size is  somewhere around  the average size  of the  packets that
          will  be transferred across the  network by your application. One
          server buffer of  whatever size you select is  allocated for each
          network task. Large network buffers will greatly effect the speed
          when moving large amounts of sequential data across the  network.
          Loading programs, using a word processor, and using a cad program
          are all  examples of the  type of applications  and use  that can
          benefit from large server buffers.

          When a workstation submits a request to a server, the server will
          work on the request until the  task is finished, or the timeslice
          allocated to  the server is  over. If the  server is set  up with
          only one network task, instead  of switching, the server will run
          the request to completion and other incoming tasks will be queued
          up behind it.  If server tasks are higher, the server can process
          several  tasks in  one active  period. One  task can  actually be
          buffered in the server while another in being processed.

          If you're  a using network  adapter with a coprocessor,  like our
          Enhanced  2mbps adapter, increasing  the number of  network tasks
          (with the Server Startup Parameters  option of NET_MGR) will give
          you  a  big  performance  boost. While  the  coprocessor  is busy
          working on one task, the  server's CPU can work on another.  This
          advantage  is multiplied  if you  have more than  one coprocessor
          equipped adapter in a  server. In fact, with three  or four 2mbps
          boards  in  a  dedicated server  and  a  corresponding number  of
          network tasks, you can achieve a continuous throughput comparable
          to that of an Ethernet network with  an equal number of stations.
          As a  plus, you've  got built-in  redundancy -- the  workstations
          connected to each adapter will  function as separate networks. If
          one  segment fails because  of cabling  or adapter  problems, the
          rest will still operate.

          SERVER REQUEST SIZE

          With NOS 4.xx a new parameter  has been added to change the  size
          of  the server  request buffer.  The REQUEST_SIZE=  is set  to 14
          bytes by default. If a request from a workstation is smaller than
          this buffer, the  server will  be able  to act  directly on  this
          request. If  the request  is larger than  the request  buffer, it
          must be buffered  by the SERVERS initial send  buffer, and system
          overhead is  increased.  The  value of 14 is  chosen specifically
          for LANtastic.   This 14 byte buffer is  large enough to tell the
          server  what type  of  operation  and how  many  bytes the  total
          operation will take.

          There are some  natural breaks in the  size of this buffer.   The
          default will handle  only the very smallest  network tasks, lock,
          unlock,  closing  a file,  etc.    At  60-70 bytes  most  network
          activity except reads and writes are processed without additional
          buffering, and from 600-2048 small reads and writes are processed
          without additional server  buffering.  This parameter  should NOT
          be set larger than the  AILANBIO Initial Send Buffer. One request
          buffer is allocated  for each user in MAX  USERS (LOGINS= command
          line switch).

          LOCK HOLD TIME

          Artisoft has discovered  that setting the  lock hold time  higher
          than 0 with LANtastic v4.1 can have an adverse effect on the time
          required  to  open  a  file.    This  parameter  should  be  used
          cautiously.  If  your  use of  the  network  entails opening  and
          closing a  large number of  files, then this parameter  should be
          set  to  a  low value.    If  your use  of  the  network involved
          accessing a large common data  file, then increasing the value of
          this parameter will help performance.

          NetBIOS and Adapter Card Memory

          With LANtastic v4.xx ethernet cards, you can increase the netbios
          packet size.   This can have an  effect on the system  speed when
          loading  programs  and  performing  large  sequential  reads  and
          writes.  Increasing  the PACKET_SIZE beyond 1500 bytes  moves the
          network out of IEEE802.3 standards  and may cause problems if you
          are using third party IEEE devices like routers or bridges.

          With  our AE-2 card,  there's an  easy way  to get  a performance
          boost  on heavily  loaded networks.  You can upgrade  the adapter
          from its default  16K configuration to a  full 64K of RAM.  It is
          fairly easy  to tell  when this  upgrade is  needed.   Start  the
          network; if  after  10  minutes  of use  lancheck  shows  network
          resource exhuasts, increasing the memory on the adapter will help
          network performance.

          Effects of Changing Buffer Sizes

          Keep in  mind that on  systems with  only two or  three stations,
          changing the  buffer configuration  may not  have any  measurable
          impact on  speed. In any case, increasing  the size and number of
          REDIR and SERVER  buffers uses memory and can  slow down programs
          running  on  the  server.  If  your  application  performance  is
          critical, consider setting  up a dedicated  server for your  most
          used application programs and printers.

          Running Applications on the Server

          To get  the best  performance out  of your  network when you  run
          applications on  a server,  you should always  use the  server as
          though  it  was a  remote  workstation:  log  in and  access  the
          application  through  network drives.  Using the  server locally,
          without logging in, places you in contention with the rest of the
          users  and network  performance can  suffer. If  you log  in, the
          server will schedule you along with the rest of the network users
          and everyone will get good performance.   This does NOT mean that
          you  should issue NET USE commands  that use a drive over itself,
          IE NET USE C: \\SERVER\C-DRIVE.  This can cause problems.   It is
          better  to create  a common  network  redirection ie  NET USE  G:
          \\SERVER\C-DRIVE  and have  ALL  workstations  (even the  server)
          access the program off of G:.

          Increasing Print Speed

          One of the slowest things about  printing to a network printer is
          spooling to the hard  disk. With LANtastic NOS 3.0 and  above you
          can easily increase the print spooling speed by spooling to a RAM
          disk.  To  change the  spooler's location,  use the  Change Spool
          Location option  of NET_MGR's  Queue Maintenance  function.   You
          must be sure  that the ramdisk is  large enough to hold  both the
          largest file printed and all print requests that may be submitted
          at one time.  The spooler will exit with an error if there is not
          enough space available on this drive.

          PRINTER BUFFERS

          As with  REDIR and  server buffers, increasing  the size  of your
          printer buffers  can improve printing  speed. If your  server has
          more  than one printer attached, increasing the number of printer
          tasks  will  allow   it  to  print  to  more   than  one  printer
          simultaneously. Artisoft has determined through testing that with
          some  applications  running  on the  server  optimum  printing is
          achieved with printer tasks set to 1 no  matter how many printers
          are  attached to  the  server.  Notably  these  applications  are
          Windows 3.x, Clipper 5.0, and  Lotus products. If you notice slow
          printing when  running one of  these programs on a  print server,
          you can  increase print speed by  setting printer tasks to  1 and
          setting cps in the printer resource to 9600.

          ARTISOFT,  Inc. makes  no warranties  as to  the completeness  or
          accuracy of this document.   LANtastic is a registered  trademark
          of ARTISOFT, Inc.   Brand names and product  names are trademarks
          or registered trademarks of  their respective companies.