IBM* LAN SERVER 4.0 WHITE PAPER
PERFORMANCE, CAPACITY ENHANCEMENTS, & TUNING TIPS
IBM LAN SYSTEMS PERFORMANCE ANALYSIS
DEPARTMENT 55LS
AUSTIN, TEXAS
MARCH 1995
(Revised)
Contents
Introduction
LAN Server 4.0 Performance Tuning Assistant
Introduction to the Tuning Assistant
HPFS 386 Cache Size Calculation
Examples of Key Parameter Calculations
Using Tuning Assistant in "What if" Mode
LAN Server 4.0 Configuration Defaults
DOS LAN Services Client Performance Considerations
LAN Server 4.0 Capacity Enhancements
LAN Server 4.0 Support of SMP
NetBIOS over TCP/IP
Design Considerations
Enhancements
Performance Characteristics
Tuning TCP/IP
Recommendation: Dual Protocol Stacks
Additional Useful Information
Reducing NetBios Broadcast Frames
DCDB Replication Performance
Upgrading from LAN Server 3.0
Considerations when RAW SMBs are disabled
DOS TCP/IP
Configuring DLS with Windows for Workgroups
Additional Tips for LAN Server 4.0 Performance
Entry vs. Advanced Server
Fixed Disk Utilization
CPU Utilization
Network Interface Cards
Network Media Utilization
Performance Benchmark Comparison
Introduction
LAN Server 4.0 includes features that allow increased capacity and
performance over LAN Server 3.0. Architectural limitations of LS 3.0 have
been addressed in LS 4.0. Parameter defaults have been increased so that a
newly installed Advanced Server supports 100 users without modifications.
Client-side caching has been added to the DOS client resulting in improved
performance. A new protocol driver that runs at ring 0 privilege along with
OS/2* NetBIOS over TCP/IP provides greater performance than the OS/2 NetBIOS
for TCP/IP used by LS 3.0. In addition, LS 4.0 provides a tool (LAN Server
4.0 Tuning Assistant) to help users tune their specific configurations for
optimum performance.
The features described above are documented in the LS 4.0 publications. It
is the intent of this paper to provide additional information, such as
design considerations and performance analysis results, from the LAN Server
Performance Analysis group.
LAN Server 4.0 Performance Tuning Assistant
Introduction to the Tuning Assistant
The Tuning Assistant was designed to satisfy the following usability and
performance objectives:
þ Provide an easy way for users to tune LS 4.0 to their configurations
þ Provide performance tuning based on users' unique situation
þ Optimize performance parameters but leave safety margin
þ Provide a tool to allow "what if" calculations
The following are general rules implemented by the Tuning Assistant during
calculations and modifications
þ Never reduce parameters below their default value
þ Never add or delete lines from any configuration file
þ Never exceed maximums such that system will not boot
þ NetBios resource requirements will be spread equally over adapters
þ LAN Server has priority if NetBios resources are overcommited
In most environments the important elements in tuning LAN Server for best
performance (in priority order) are the following:
þ Configure the largest HPFS386 cache possible
þ Provide a sufficient number of NUMREQBUF
þ Provide a sufficient number of Commands ('x2' parameter in the
'netx' line of the IBMLAN.INI)
þ Enough adapters (and NETBIOS resources) for number of users
þ Provide a sufficient number of NUMBIGBUF(Entry Server or print
spooling only)
þ Reserve sufficient memory for GUI if it will be used frequently and
delay for swapping is undesirable
HPFS386 Cache Size Calculation
The HPFS386 cache size calculation involves determining all the uses of
memory in the system and assigning the remainder to the cache. Listed
below are the factors used by the Tuning Assistant in the calculation of
the HPFS386 cache size of a system with a memory size of 32.0 MB.
Memory Allocation (MB):
- OS/2 base 2.8
- Spooler 0.7
- MPTS base 0.6
- Memory per adapter 0.2
- LAN Server base 3.5
- IBMLAN.INI additional 1.5
- Heap reserve 1.0
- Cache mgmt (64 bytes/1K 1.3
bytes for memory in excess
of 12M)
- Reserve for other apps 1.0
- Safety margin (5 percent) 1.6
- Memory assigned to HPFS386 cache 17.8
TOTAL 32.0 MB
The entries for OS/2, MPTS and LAN Server base are the same as found in the
Memory Estimating Worksheets in the Network Administrators References
Volume 1, Appendix A.
The "IBMLAN.INI additional" entry is for increases made to certain
parameters by the Tuning Assistant.
The Heap reserve memory is set to 1 MB; this memory is used by LAN Server
for internal file system control needs such as open file handle tables,
search handles, filename parsing, etc. This memory is not assigned to
the Heap parameter but merely set aside for availability.
The cache management entry is necessary when dealing with large cache sizes
as it becomes a significant amount of memory. The formula is applied
after subtracting 12 MB from the system memory size since at least 12 MB
of memory is always needed and is never available for cache.
By default, 1 MB is always reserved for other user applications to be run
on the server. This is a parameter used by the Tuning Assistant to
provide the administrator with a significant input to the cache size
calculation. Note: The administrator should determine the memory
requirements of any application that is to run concurrently with LAN
Server and provide that value in the 'Application Reserve Memory' entry
field of Tuning Assistant.
An important example of an additional user application on the server is
the new LS 4.0 Graphical user Interface (GUI). If the administrator will
regularly use the GUI administration feature, at least 5 MB should be
entered for this parameter to provide good performance of the GUI. If
this amount of memory is not available, significant swapping will occur
when the GUI is started. If only occasional use of the GUI on the server
is expected, the recommendation is to leave this parameter at 1 MB and use
the additional 4 MB system memory for HPFS386 cache.
Examples of Key Parameter Calculations
NUMREQBUF (IBMLAN.INI)
Optimum number is 2 to 3 per "active" user. Since NUMREQBUF locks the
memory from other processes, we want to be efficient. Also, since most
uses of NUMREQBUF require a corresponding command it is wasteful to
allocate more NUMREQBUF than commands. Only 250 commands per adapter are
configured by Tuning Assistant, thus only 250 NUMREQBUFs per adapter will
be configured.
Calculation: 2.2 times MAXUSERS with maximum of 250 per adapter.
Special Considerations: Memory used by NUMREQBUF is calculated in Tuning
Assistant using a hard coded value of 4096 bytes for each request
buffer. If the user wants to change SIZREQBUF from 4096 to 2048,
then the calculated HPFS386 cache size can be increased by
(NUMREQBUF * SIZREQBUF) /2.
A parameter related to NUMREQBUF is the USEALLMEM parameter in the
Requester section of the IBMLAN.INI. This parameter allows request
buffers to be defined in the memory above 16 MB. If no network interface
cards (NICs) are limited to 24-bit direct memory access (DMA), and more
than 16 MB of RAM is installed in the machine, set this parameter to
'YES'.
NUMBIGBUF (IBMLAN.INI)
NUMBIGBUF are used only by the ring 3 (Entry) Server when files are
accessed on a FAT or HPFS file system or the Printer Spooler is accessed.
Because better performance can be obtained by using all available memory
for HPFS386 cache, NUMBIGBUF will not be increased if the LAN Server
Advanced package is installed.
Calculation: If Advanced Server, NUMBIGBUF = 12 (default). If Entry
Server, NUMBIGBUF increases to a maximum of 80 as MAXUSERS increases.
Commands (IBMLAN.INI)
For optimum performance commands also needs to be 2 to 3 per "active" user
because of its close relationship with NUMREQBUF. Obviously, if 250 users
are logged on through one adapter, each user will not have 2 to 3 commands
always available and performance will be less than optimum. The
IBMLAN.INI parameter that specifies commands is 'x2' in the 'netx'
statement.
Calculation: 2.2 times MAXUSERS with maximum of 250 per adapter
Special Considerations: Commands (NCBS in Protocol.ini) are NetBios
resources and must be shared with other NetBios applications like DB2*,
Lotus Notes**, etc. If a user specifies a MAXUSERS >=114, commands of
250 will be set for LAN Servers' net1 line, leaving only 4 commands for
other NetBios applications. Users should manually reduce commands in the
net1 line to allow the other applications more NCBS resources if required.
Maxusers (IBMLAN.INI)
Calculation: # DOS/Windows** users + # OS/2 users + # additional servers
(if Domain Controller)
Maxshares (IBMLAN.INI)
Calculation: Number home dirs + # aliases + (3 * number shared apps).
Maxconnections (IBMLAN.INI)
Calculation: (MAXUSERS + # additional servers) * 4
Special considerations: Advanced Server maintains its own set of connection
resources; this parameter pertains only to resources shared by the ring 3
(Entry) Server such as print aliases. This is also true for the following
parameters which are not changed for Advanced Server (if HPFS only):
MAXLOCKS, MAXOPENS.
Maxsearches (IBMLAN.INI)
Calculation: The Tuning Assistant sets MAXSEARCHES = 700
Special considerations: Advanced Server maintains its own set of search
resources; this parameter pertains only to searches done by the ring 3
(Entry) Server. This value was chosen to provide ample search memory for
the ring 3 (Entry) Server.
Sessions (PROTOCOL.INI)
Calculation: DOS/WIN requesters + OS/2 requesters + # additional servers
(if DC) + Lotus Notes requesters + DB2 requesters + User
logged on server + other NetBios requirements
NCBS Calculation (PROTOCOL.INI)
For optimum performance NCBS also need to be 2 to 3 per "active" LAN Server
user plus the NetBios commands needed by other NetBios applications.
Calculation: 2.2 times MAXSESSIONS + other NetBios reqmts up to a maximum
of 254 per adapter
Special considerations: NCBS in PROTOCOL.INI are shared with other NetBios
applications like DB2, Lotus Notes, etc. LAN Server will use a maximum of
250 of the 254.
Using Tuning Assistant in "What if" Mode
In response to feedback from Beta users the "What if" mode was added,
although it is not described in the product documentation. This is the
capability to run Tuning Assistant calculations on a machine other than
where LAN server 4.0 is installed. This will allow a user to provide system
configuration information to Tuning Assistant and create tuned
configuration files for use on other machines. IBM is interested in your
experience with Tuning Assistant. Please post your comments and any
operational concerns to the LS40 CFORUM and someone will respond.
The Tuning Assistant's filename is LS40TUNE.EXE; it is located in the
IBMLAN directory. The additional parameters which can be used in the
command line launch of Tuning Assistant are as follows:
/D:DOMAIN1 - Domain name (has no effect on calculations)
/S:SERVER1 - Server name (has no effect on calculations)
/T:DC(or AS) - Type: Domain controller or Additional server
/P:Entry(or Advanced) Package - Entry or Advanced version
/M:XX - System Memory in MB
/A:N - Number of network interface cards(adapters)
/U - User supplied files (CONFIG.SYS, IBMLAN.INI, PROTOCOL.INI,
and HPFS386.INI)
NOTE: /T is always required whenever /U is specified if running on
a system with no server installed.
Example 1
LS40TUNE /D:DOMAIN1 /S:SERVER1 /T:DC /P:ADVANCED /M:32
/A:2 /U
To run Tuning Assistant this way, all four of the Advanced version
configuration files must be located in the current subdirectory with
LS40TUNE. This will run on a machine with or without a server installed.
The command line values will take precedence over any actual system version
of these parameters. If the "Apply" pushbutton is chosen the user supplied
files will be changed and no backup files will be made.
Example 2
LS40TUNE /M:32 /A:2
This will run only on a machine with a server installed. The command line
values will take precedence over actual system versions of these
parameters. This example could be useful to look at the effects of more
system memory or network interface cards on the tuning calculations. When
the "Apply" pushbutton is chosen the system configuration files will be
changed and backup files will be copied into the \IBMLAN\BACKUP
subdirectory with names like IBMLAN.001, PROTOCOL.001, etc. All files that
are updated when the Tuning Assistant calculation is 'applied' will have
the same suffix.
Warning: The "What if" feature is useful in examining the logic of the
Tuning Assistant, but you should be careful when creating actual
configuration files for use on systems other than the one on which
the tool was executed.
LAN Server 4.0 Configuration Defaults
The Advanced version of LS 4.0 may be used in larger configurations than
previous versions. Therefore, default values of a number of parameters have
been increased. The objective is to allow many users to run LS 4.0 out-of-box
with little or no customized tuning. The Advanced Server will support 100
users in typical environments, however, running the Tuning Assistant may
provide an additional performance improvement for some customers. Some
changes to the Entry Server and Peer Services were also made. A summary
follows:
IBMLAN.INI ADVANCED ENTRY PEER
PARAMETERS SERVER SERVER SERVICES
LS 3.0 LS 4.0 LS 3.0 LS 4.0 LS 3.0 LS 4.0
| maxopens | 576 | 256 | 576 | 160 | 576 | 128 |
|--------------------------------------------------------------------
| maxsearches | 50 | 350 | 50 | 150 | 50 | 50 |
---------------------------------------------------------------------
| numbigbuf | 12 | 12 | 12 | 6 | 12 | 4 |
---------------------------------------------------------------------
| numreqbuf | 36 | 250 | 36 | 48 | 36 | 10 |
---------------------------------------------------------------------
| maxshares | 16 | 192 | 16 | 64 | 16 | 16 |
---------------------------------------------------------------------
| maxusers | 32 | 101 | 32 | 32 | 5 | 5 |
---------------------------------------------------------------------
| maxconnections | 128 | 300 | 128 | 128 | 26 | 26 |
---------------------------------------------------------------------
| x1(in net1) | 32 | 102 | 32 | 34 | 32 | 34 |
---------------------------------------------------------------------
| x2(in net1) | 50 | 175 | 50 | 70 | 50 | 70 |
---------------------------------------------------------------------
---------------------------------------------------------------------
|------------- ---------------|
| PROTOCOL.INI| SAME FOR ALL |
| PARAMETERS | VERSIONS |
-------------------------------
| | LS 3.0| LS 4.0|
-------------------------------
| sessions | 40 | 130 |
-------------------------------
| ncbs | 95 | 225 |
-------------------------------
HPFS386 Cache Defaults
The HPFS386 cachesize was specified in the IFS line in CONFIG.SYS in LS
3.0. For LS 4.0 it is specified in the \IBM386FS\HPFS386.INI file with a
line reading "cachesize = xxxx" in the FILESYSTEM section.
The algorithm for determining the default HPFS386 cache size has also
changed. Previously the cache size was set at 20 percent of the remaining
memory after OS/2 was started. This gave a cache size of 2.9 MB on a 16 MB
system. This formula will still be used as long as there is less than 20 MB
of memory in the system. If the system has at least 20 MB of memory and the
user has indicated that the server can use memory above 16 MB for cache,
the default cache size will be 60 percent of remaining memory after OS/2
has started. This will yield a cache size of around 18 MB on a 32MB system.
This will enable LS 4.0 to provide excellent performance on most systems
without any tuning.
As with earlier releases of LAN Server, the USEALLMEM parameter defaults
to 'NO'. This restricts access to memory above 16 MB. If no network
interface cards (NICs) or disk adapters are limited to 24-bit direct memory
access (DMA), and more than 16 MB of RAM is installed, this parameter
should be set to 'YES'. This parameter used to be in the CONFIG.SYS file
but is now in the FILESYSTEM section of the HPFS386.INI file.
DOS LAN Services Client Performance Considerations
OS/2 Lan Server 4.0 comes with DOS Lan Services (DLS) clients. DLS clients
offer substantial performance improvements over the DOS Lan Requester (DLR)
clients provided with LAN Server 3.0. Significant performance benefit is
realized through the implementation of client-side caching algorithms. In
brief, client-side caching offers local caching, reducing requests to the
server, thereby increasing overall system performance.
Client-side caching is enabled by default with DLS clients. This means that
the AUTOCACHE parameter is set to YES in the NETWORK.INI file (since this
a default, it is not specifically written into NETWORK.INI during
installation). With AUTOCACHE=YES, the DLS client will allocate big buffers
in extended memory(XMS). Each big buffer is 8K in size. The number of big
buffers is calculated by the system and is dependent on the amount of XMS
available (up to a maximum of 30 big buffers). If a machine does NOT have
any XMS the AUTOCACHE parameter is effectively ignored. If you want to
configure big buffers on a DLS client that has no XMS, set the following
parameters in the NETWORK.INI file:
1. AUTOCACHE=NO
2. SIZBIGBUF=xxxx (in bytes)
3. NUMBIGBUF=xx (integer)
This will allocate big buffers on the client that can be used for large data
transfers. However, these buffers will be put in Upper and/or Conventional
Memory, reducing available memory for applications.
Another parameter of importance with DLS clients is the WORK BUFFERS.
These are the buffers that are used on the requester to process an
application's request for data. The default values for WORK BUFFERS on the
DLS client are as follows (also set in NETWORK.INI).
1. SIZWORKBUF=1024
2. NUMWORKBUF=2
The above default values are the generally recommended values for the best
system performance. If you are unable to use the AUTOCACHE option, you may
want to experiment with these two parameters for possible improvements in
your environment.
LAN Server 4.0 Capacity Enhancements
As the number of workstations connected to LAN Server 3.0 grew into the
hundreds in some installations, an architectural limitation was discovered
which has been addressed in LAN Server 4.0. Specifically, a data structure
design was limiting the number of request buffers(NUMREQBUF) which could be
configured to a maximum of around 350. In large installations this could
cause a performance degradation. LAN Server's new design provides future
extensibility by allowing the value of NUMREQBUF to be as large as 2000.
The current recommended value for NUMREQBUF is 2.2 per user up to a
maximum of 250 for each adapter, or 1000 if four adapters are in the system.
LAN Server 4.0 Support of SMP
LAN Server 4.0 has been tested with and shown to support symmetric
multi-processor(SMP) machines running under OS/2 for SMP. LAN Server 4.0
Advanced does not gain additional performance benefits from SMP machines.
Its architecture has been optimized to the point where most requests are
processed "on interrupt" when received from the network component. The
queuing time for a request to be processed is usually extremely short since
there are rarely instances when a file/print server's CPU approaches 100
percent utilization. Under these conditions, it would not be expected that
an additional CPU would improve response time to the requester. This design
provides industry leading performance as evidenced by the LANQuest** report
of October 1994. See page 25 for more information.
There are some situations in which LAN Server 4.0 support of SMP does lead
to an improvement in total system throughput performance. Since OS/2 is a
multi-tasking operating system, other applications can run in the same
machine as LAN Server. For other applications which make extensive use of
the CPU(e.g. Lotus Notes, etc.), additional processors may make sense.
Whenever the CPU workload approaches 100 percent, the additional processor
can make a significant difference in the system throughput. LAN Server 4.0
Advanced accommodates the use of the additional processor unless its own
workload is unusually high in which case it takes precedence over other
applications. LAN Server 4.0 Entry runs with the same privilege as other
OS/2 applications and does not take precedence in an SMP environment.
NETBIOS Over TCP/IP
Design Considerations
NetBIOS over TCP/IP is an implementation of NetBIOS that has been
specifically designed to operate with IBM TCP/IP. It enables a
workstation to be geographically isolated from its domain yet communicate
with it transparently.
NetBIOS over TCP/IP is an implementation of the Request for Comments (RFCs)
1001/1002 standards which describe how to enable NetBIOS applications over
TCP/IP. It is a B-node, or Broadcast node implementation with routing
extensions. A broadcast node uses broadcasting to exchange information
between hosts. The routing extensions allow nodes to span subnets through
IP routers. These extensions plus the remote name cache discussed below
simplify the configuration of RFC 1001/1002 NetBIOS nodes into TCP/IP
environments.
NetBIOS over TCP/IP uses an expanded syntax for NetBIOS names that is
transparent to NetBIOS applications. The Local NetBIOS Name Scope String
is appended to the NetBIOS name creating an expanded name that has the
effect of limiting the scope of a NetBIOS name. Two RFC-compliant NetBIOS
nodes can communicate only if they have the same Local NetBIOS Name Scope.
The Local NetBIOS Name Scope string is defined by the LOCALSCOPE parameter
in the TCPBEUI section of the PROTOCOL.INI.
NetBIOS over TCP/IP supports only 1 logical netbios adapter and should
therefore be added to only 1 network interface card during the installation
/configuration process. However, if TCP/IP is installed on multiple
adapters, NetBIOS over TCP/IP will make use of those adapters.
TCPBEUI is IBM's high performance, ring zero protocol driver which maps
NetBIOS API calls into the TCP/IP protocol. NetBIOS over TCP/IP contains
enhancements over the RFC 1001/1002 standards which improve system
performance by decreasing broadcast storms, and expanding communications
over routers and bridges. These enhancements, described in the next
section, are transparent to NetBIOS applications and do not interfere
with other B-node implementations that lack similar functions.
Enhancements
Three of the enhancements to NetBIOS over TCP/IP are in the form of routing
extensions. These extensions allow communication between networks and over
IP routers and bridges. These extensions are:
1. The broadcast file. A broadcast file contains a list of host names, host
addresses, or directed broadcast addresses. It is read at startup
and each valid address is added to the set of destination addresses
for broadcast packets. Remote nodes included in the broadcast file
are then treated as if they were on the local network. Use of a
broadcast file has the effect of extending a node's broadcast domain
to its own subnet plus any other subnets listed in the broadcast
file. A maximum of 32 broadcast file entries are supported, each
of which could include additional subnets, thus extending the node's
broadcast domain.
2. The names file. A names file consists of NetBIOS name and IP
address pairs. NetBIOS over TCP/IP will conduct a prefix search of
the names file before broadcasting on the network. The prefix match
succeeds if the entry in the names file matches the given name, up
to the length of the entry. The first match is used, therefore, the
order in which NetBIOS names are listed in the names file is
important.
To enable this routing extension, set the NAMESFILE parameter in the
TCPBEUI section of the PROTOCOL.INI to a nonzero integer that
represents the number of names file entries.
3. The Domain Name Server (DNS). A network administrator can maintain
NetBIOS name and IP address pairs in a DNS. If a name query fails
NetBIOS over TCP/IP can append the NetBIOS Domain Scope String to
the encoded NetBIOS name and issue a request to the DNS to look up
an IP address for that NetBIOS name.
The Domain Scope String is defined by the PROTOCOL.INI parameter
DOMAINSCOPE.
Another enhancement NetBIOS over TCP/IP provides is a cache for storing
remote names that have been discovered. This cache is enabled by setting
the NAMECACHE parameter in the TCPBEUI section of the PROTOCOL.INI to a
nonzero integer that represents the number of names stored in the directory
(NAMECACHE=xx).
The information in the remote names cache (or directory) is also stored on
disk and periodically updated. When the system is restarted, this
information can be preloaded into the cache at bootup time. Preloading
can reduce the amount of broadcast frames on the network since NetBIOS
will not have to rediscover names for remote names. To preload the remote
names cache, set PRELOADCACHE=YES in the TCPBEUI section of the
PROTOCOL.INI.
NOTE: When NetBIOS over TCP/IP is searching for a name, the name cache is
checked first, then the names file, the broadcast file, and finally the
Domain Name Server.
Recommendation: When running NetBIOS over TCP/IP in a Wide Area
Network (WAN), turn name caching on at the server (e.g. NAMECACHE=100).
Performance Characteristics of NetBIOS over TCP/IP
The performance difference between NetBIOS over TCP/IP and NetBEUI can
range widely depending on the environment. Some environmental factors
that can affect performance are the type of client (OS/2 or DOS), the
server CPU workload, the type of network operations being performed, the
network media, network congestion, and communication line speeds. We've
observed the performance of NetBIOS over TCP/IP being anywhere from 10%
slower to as much as 4 times slower than NetBEUI.
One of the environments in which we conducted performance tests was a
medium- sized Local Area Network on 16Mbps Token Ring with no WAN
connections. We ran a set of industry standard business applications on
OS/2 NetBIOS over TCP/IP clients and again on OS/2 NetBEUI clients. In
this environment, NetBIOS over TCP/IP was 20% slower than NetBEUI. The
performance of DOS NetBIOS over TCP/IP clients was significantly less
than that of the OS/2 clients.
Database applications generally use small records when accessing shared
databases residing on the server. Often these small records are retrieved
from the file system cache with no physical disk access being required.
The performance of this type of application on NetBIOS over TCP/IP may
be noticeably slower than if the application were run using NetBEUI.
However, if the number of database accesses of this type in performing a
typical operation is in the order of hundreds, not thousands, the user may
not notice a difference in performance in the two protocols.
It may be necessary to periodically update client applications or other
files by copying them from the server disk. DCDB replication from a
domain controller to a remote additional server also generates I/O
operations sometimes known as file transfers. This type of file I/O
activity over a network will show little or no performance difference
between NetBEUI and NetBIOS over TCP/IP due to protocol characteristics.
One should be aware, however, that most WAN connections today are made
over relatively low speed communication lines when compared with a LAN
speed of 4 to 16 Mbps. File transfer operations over WAN communication
lines will probably be slower than over LANs but most likely not due to
the network protocol.
Tuning TCP/IP
If you're using NetBIOS over TCP/IP in a Local Area Network environment,
file transfer performance might be improved by increasing the maximum
transmissible unit (MTU) size. We have seen up to a 20 percent increase
in performance of large file transfers by using an 8KB packet instead of
the default 1500 bytes. The default of 1500 was chosen because of
ethernet's packet size limitation and prevalence in TCP/IP environments.
The MTU size can be changed with the IFCONFIG command in TCP/IP's
SETUP.CMD. Set the MTU size to the desired packet size plus 40 bytes, the
maximum TCP/IP header size. The desired packet size should be a multiple
of 2048.
Your network adapter must be configured to support transmission of buffers
that are at least the size specified for the MTU. On an IBM 16/4 Token Ring adapter,
this would be accomplished by setting the XMITBUFSIZE parameter in the
Token Ring section of the PROTOCOL.INI file. Check your network interface
card documentation for information on configuring your adapter.
Recommendation: Dual Protocol Stack
Because there may be a performance difference in a particular environment,
it is recommended to configure and use NetBEUI in the Local Area Network
(LAN) environment, and NetBIOS over TCP/IP in the Wide Area Network (WAN)
environment. The Multi-Protocol Transport Services (MPTS) shipped with LAN
Server 4.0 provides the capability of configuring your LAN workstation or
server with both NetBEUI and NetBIOS over TCP/IP on the same network
interface card.
The dual protocol stack can be configured through the LAN Server
installation/configuration program. When selecting protocols, install
logical adapter 0 with NetBEUI and logical adapter 1 with TCP/IP and
NetBIOS over TCP/IP. This dual protocol stack configuration allows local
sessions to continue running with NetBEUI performance while also providing
Wide Area Network connectivity with NetBIOS for TCP/IP.
Additional Useful Information
Reducing NetBIOS Broadcast Frames
A key concern with many NetBIOS users is the amount of broadcast traffic
that occurs on the network. Broadcasts are used to communicate between
nodes. Broadcast storms can slow network performance and overwhelm routers.
Use of the Remote Name Directory (RND) function can help to minimize this
broadcasting by sending frames to specific nodes when possible.
When using RND, the local station caches the node addresses of remote names
that it has located. Any messages sent to that remote name after the node
address has been saved is sent directly to that node rather than broadcast
to all nodes.
The RND function in LAN Server 4.0 has been extended to include datagrams.
RND stores only unique names and no group names, so if an application uses
mostly group names for sending datagrams, RND should not be used. Another
enhancement to the RND function is that the maximum number of directory
entries has been increased from 255 to 2000 when running on OS/2 2.0 or
greater.
The parameter RNDOPTION in the NETBEUI section of the PROTOCOL.INI
specifies whether RND is turned on or off. Set this parameter to 1 to
enable use of the RND function. If RNDOPTION is chosen, make sure that
DATAGRAMPACKETS in the NETBEUI section is greater than 2. A related
parameter, also found in the NETBEUI section, is NAMECACHE. This parameter
specifies the size of the remote name directory. This parameter defaults
to 1000 entries.
DCDB Replication Performance
Changes to the DCDB Replicator service for LS 4.0 have yielded substantial
performance improvements. In some configurations, users may see up to an
80 percent increase in performance over the LS 3.0 DCDB Replicator service.
Upgrading from LAN Server 3.0
Upgrading from LAN Server 3.0 to LAN Server 4.0 will cause parameters in
the PROTOCOL.INI file to be set to the LS 4.0 default values. This may
cause performance problems in previously tuned servers. Users who have
fine-tuned their PROTOCOL.INI for LS 3.0 should be aware that they may
need to make the same changes for LS 4.0.
Considerations when RAW SMBs are disabled
The multiplex read and write SMB protocols are used if the RAW SMB protocol
is disabled. These protocols divide data transfers into buffer-size chunks
(sizworkbuf) and chain them together to satisfy large read or write
requests.
A parameter that affects performance when working in multiplex mode is
PIGGYBACKACK in the NetBEUI section of the PROTOCOL.INI file. This
parameter specifies whether NetBIOS sends and receives acknowledgements
piggybacked with incoming data. When used with RAW SMBs, piggybackacks
improve performance. However, users that attempt to use piggybackacks with
multiplex SMBs may see performance degrade by up to 3 times for large file
transfers.
Note: The RAW SMB protocol is disabled on a server when srvheuristic 19 in
the IBMLAN.INI file is set to 0 (default=1). The RAW SMB protocol on an
OS/2 client is disabled when IBMLAN.INI wrkheuristic 11 is set to 0
(default=1) and wrkheuristics 14 and 15 are set to 1 (default=1).
DOS TCP/IP
The LAN Server Performance Team has tested a number of vendor TCP/IP
products for DOS. These include Network Telesystems, Wollongong, and FTP
TCP/IP offerings. In many cases, these performed considerably better than
the IBM TCP/IP protocol stack shipped with LAN Server 4.0. The Network
Telesystems product, in particular, showed significant throughput
improvement.
While IBM continues to refine their DOS TCP/IP offering, the performance
content of each of the OEM products reviewed may provide a near-term
solution for running DOS clients in a TCP/IP environment.
In addition to the TCP/IP protocol stack, each of the vendor products
includes the normal TCP/IP applications such as FTP, mail, SNMP, etc.
Configuring DOS LAN Services with Windows for Workgroups
You can install both Windows for Workgroups and DOS LAN Services on the
same workstation. However, you cannot use the network function of Windows
for Workgroups with this configuration. To run DOS LAN Services and
Windows for Workgroups on the same workstation, use the following
procedure:
1. Install Windows for Workgroups
2. Install DOS LAN Service
3. In the WINDOWS\SYSTEM directory, rename the following files:
From: To:
VNETSUP.386 VNETSUP.WFW
VREDIR.386 VREDIR.WFW
NETAPI.DLL NETAPI.WFW
PMSPL.DLL PMSPL.WFW
4. In the CONFIG.SYS file, REM out the following line:
'DEVICE=C:\WINDOWS\IFSHELP.SYS'
5. In the Windows SYSTEM.INI file, under the '386enh' section,
change the line that contains the 'network=' statement to the
following: 'network=vnetbios.386,vnetsup.386,vredir.386'
The fix for APAR IC08963 makes the same changes, so you can use the APAR
fix if you do not want to change the CONFIG.SYS and SYSTEM.INI files
manually.
Additional Tips for LAN Server 4.0 Performance
A number of the major factors affecting performance of LAN Server 4.0 are
reviewed in the following sections. Although a few parameters are discussed,
most of the tips are aimed at getting you to think about your particular
environment in relation to LAN Server's system resources.
Because there will always be a bottleneck in any computer system, the
objective of performance tuning is to remove the current bottleneck.
Hopefully the resulting system performance has its new bottleneck at an
operating point outside normal operating conditions.
Entry vs. Advanced Server
If your LAN Server is to share file, applications, or printers for less
than 80 users, the Entry Server will fit your needs with very good
performance. The LANQuest report described on page 22 contains a
comparison of Entry vs. Advanced Server performance. A subsequent upgrade
to Advanced Server is available with minimum impact to your business.
If your immediate requirements are for high performance and high capacity,
you will want the Advanced Server. To gain the performance advantage of
the Advanced version, your applications and data files must reside on an
HPFS386 partition, not on a FAT partition. Neither OS/2 2.1 nor LS 4.0
must be installed on an HPFS386 partition because accesses to system
software are infrequent after initial loading.
Fixed Disk Utilization
The disk subsystem, an electromechanical device, can often be the system
bottleneck even though the system provides a lot of memory for caching
files. If you have observed that your fixed disk activity indicator (the
little light that flashes when the hard disk is in use) is on more than it
is off for long periods of time, you probably have a disk bottleneck.
Your options for improving performance include:
- Distribute the disk-intensive workload from a single physical disk
drive to multiple disk drives, enabling concurrent disk seeks and
read/writes.
- Off-load some users, files, or applications to another server.
- Install the Fault Tolerance feature of LAN Server to enable disk
mirroring. This not only protects your data by backing up your disk
but also improves performance since the additional disk drive will
also be used to read data (split reads).
- Adding fixed disks and striping data across them (RAID architecture)
will sometimes improve performance as well as enhance data integrity
in an environment where data is predominantly looked up (read)
without a subsequent update (write), for example, databases used
for price lookup, part number information, etc.
CPU Utilization
Server performance can degrade when the computer (CPU)'s ability to process
incoming instructions is overtaxed. If there are many users (usually
hundreds) with high interaction rates to the server, a CPU performance
bottleneck may occur (the Advanced server CPU efficiency is several times
greater than the Entry server). You may see a lot of fixed-disk activity
and suspect the disk subsystem, but this may be lazy-write activity, which
is not necessarily the system bottleneck. To check CPU utilization you
can use System Performance Monitor/2 or LAN NetView* Monitor for a
detailed analysis.
To get a rough idea of how your server uses the CPU, start the Pulse
applet from the OS/2 Workplace Shell* Productivity folder and observe its
display during a heavy server workload period. If the CPU utilization
level is 80 percent or greater for much of the time, performance is being
impacted by the CPU's ability to satisfy its workload demands. Replacing
standard network interface cards (NICs) with busmaster NICs will provide
additional CPU power and usually improve server performance. Another
remedy is to offload some of the users, files, applications, or functions
(e.g., domain controller or print server) to another server or to upgrade
to a more powerful hardware system.
Network Interface Cards (NICs)
Let's assume that your fixed disk activity is not excessive and that your
CPU utilization is generally less than 30 to 40 percent, but you still
feel that your server could respond more quickly. Your network interface
card (NIC) is analogous to a nozzle which physically limits the amount of
traffic flowing to/from the server. Depending on the number of users,
speed of the client machines, type of data transactions, etc., server
performance can be NIC-limited. NICs come in 8-bit, 16-bit and 32-bit bus
widths. Some 32-bit NICs are busmasters, which means they can handle most
data transfers with their built-in processors, relieving the server CPU of
this task.
You can improve an NIC-limited condition by changing to a faster NIC and/or
adding additional NICs to your server. As you add additional NICs, your
server CPU utilization will increase as the server will be busier than
before servicing the additional traffic coming through the NICs (nozzles).
If you add busmaster NICs, the increase in server CPU utilization will be
less significant, as you might expect. LS 4.0 will automatically load-
balance sessions across all NICs when you initiate a session. When using
standard 16/4 token-ring NICs, we recommend that you use a 16 KB shared
RAM size for best performance and memory utilization.
Both LAN Server versions 3.0 and 4.0 now support more OEM NICs than the
initial release of LAN Server 3.0. You may obtain the current lists of
supported NICs from CompuServe** with the following selections:
1. GO IBM
2. Technical Service and Support
3. IBM OS/2 Forums
4. OS/2 Developer 2 Forum (Browse)
5. LAN Server Library
Network Media Utilization
The physical media over which network traffic flows has a finite capacity.
The Ethernet bandwidth limit today is usually 10 megabits per sec (Mbps);
token rings today are running at 4 Mbps or 16 Mbps. It is quite possible
that with powerful servers and hundreds of clients, LANs can almost
saturate the physical media providing interconnection. This is much more
likely to occur in Ethernet networks due to the broadcast/collision
detection/re-broadcast nature of that architecture.
In large networks interconnecting many clients and servers, the level of
network traffic on the wire can impact token- ring network performance.
A simple (but not always viable) remedy is to change your network topology. You could add
NICs to your server, and separate and isolate clients into LAN segments so
that all network traffic is not passing through all machines. The net
effect is that the server with two Ethernet NICs now has a greater
potential bandwidth (20 Mbps) plus a lower collision level on each of the
two segments than on a single Ethernet segment.
This solution is not viable if the machines on the two isolated segments
must communicate, since LS 4.0 does not internally route the NetBIOS
protocol. More sophisticated ways to reduce network utilization include
using the traditional backbone rings and bridges plus the new intelligent
switches, hubs, and routers now becoming available.
Performance Benchmark Comparison
In October 1994 LANQuest Labs published a Performance Benchmark
Comparison Report assessing the performance of LAN Server 4.0 Advanced and
Entry, Windows NT** Server 3.5, and NetWare** 4.02. The results of this
benchmarking showed that LAN Server 4.0 Advanced was 38% faster than
Windows NT Server and 11% faster than NetWare. For copies of this report,
call 1-800-IBM-4FAX and request document 2014.
Trademarks denoted by an asterisk (*) are IBM trademarks or registered
trademarks of the IBM Corporation in the United States and/or other
countries: IBM, OS/2, DB2, NetView, Workplace Shell
Trademarks denoted by a double asterisk (**) are registered trademarks of
their respective companies.