Tuning the Operating System and Platform

5 Tuning the Operating System and Platform

This chapter discusses tuning the operating system (OS) for optimum performance. It discusses the following topics:

Server Scaling
Solaris 10 Platform-Specific Tuning Information
Tuning for the Solaris OS
Tuning for Solaris on x86
Tuning for Linux platforms
Tuning UltraSPARC CMT-Based Systems

Server Scaling

This section provides recommendations for optimal performance scaling server for the following server subsystems:

Processors
Memory
Disk Space
Networking
UDP Buffer Sizes

Processors

The GlassFish Server automatically takes advantage of multiple CPUs. In general, the effectiveness of multiple CPUs varies with the operating system and the workload, but more processors will generally improve dynamic content performance.

Static content involves mostly input/output (I/O) rather than CPU activity. If the server is tuned properly, increasing primary memory will increase its content caching and thus increase the relative amount of time it spends in I/O versus CPU activity. Studies have shown that doubling the number of CPUs increases servlet performance by 50 to 80 percent.

Memory

See the section Hardware and Software Requirements in the GlassFish Server Release Notes for specific memory recommendations for each supported operating system.

Disk Space

It is best to have enough disk space for the OS, document tree, and log files. In most cases 2GB total is sufficient.

Put the OS, swap/paging file, GlassFish Server logs, and document tree each on separate hard drives. This way, if the log files fill up the log drive, the OS does not suffer. Also, its easy to tell if the OS paging file is causing drive activity, for example.

OS vendors generally provide specific recommendations for how much swap or paging space to allocate. Based on Oracle testing, GlassFish Server performs best with swap space equal to RAM, plus enough to map the document tree.

Networking

To determine the bandwidth the application needs, determine the following values:

The number of peak concurrent users (N _peak) the server needs to handle.
The average request size on your site, r. The average request can include multiple documents. When in doubt, use the home page and all its associated files and graphics.
Decide how long, t, the average user will be willing to wait for a document at peak utilization.

Then, the bandwidth required is:

N_peakr / t

For example, to support a peak of 50 users with an average document size of 24 Kbytes, and transferring each document in an average of 5 seconds, requires 240 Kbytes (1920 Kbit/s). So the site needs two T1 lines (each 1544 Kbit/s). This bandwidth also allows some overhead for growth.

The server’s network interface card must support more than the WAN to which it is connected. For example, if you have up to three T1 lines, you can get by with a 10BaseT interface. Up to a T3 line (45 Mbit/s), you can use 100BaseT. But if you have more than 50 Mbit/s of WAN bandwidth, consider configuring multiple 100BaseT interfaces, or look at Gigabit Ethernet technology.

UDP Buffer Sizes

GlassFish Server uses User Datagram Protocol (UDP) for the transmission of multicast messages to GlassFish Server instances in a cluster. For peak performance from a GlassFish Server cluster that uses UDP multicast, limit the need to retransmit UDP messages. To limit the need to retransmit UDP messages, set the size of the UDP buffer to avoid excessive UDP datagram loss.

To Determine an Optimal UDP Buffer Size

The size of UDP buffer that is required to prevent excessive UDP datagram loss depends on many factors, such as:

The number of instances in the cluster
The number of instances on each host
The number of processors
The amount of memory
The speed of the hard disk for virtual memory

If only one instance is running on each host in your cluster, the default UDP buffer size should suffice. If several instances are running on each host, determine whether the UDP buffer is large enough by testing for the loss of UDP packets.

Note:

On Linux systems, the default UDP buffer size might be insufficient even if only one instance is running on each host. In this situation, set the UDP buffer size as explained in To Set the UDP Buffer Size on Linux Systems.

Ensure that no GlassFish Server clusters are running.
If necessary, stop any running clusters as explained in "To Stop a Cluster" in GlassFish Server Open Source Edition High Availability Administration Guide.
Determine the absolute number of lost UDP packets when no clusters are running.
How you determine the number of lost packets depends on the operating system. For example:
- On Linux systems, use the netstat -su command and look for the packet receive errors count in the Udp section.
- On AIX systems, use the netstat -s command and look for the fragments dropped (dup or out of space) count in the ip section.
Start all the clusters that are configured for your installation of GlassFish Server.
Start each cluster as explained in "To Start a Cluster" in GlassFish Server Open Source Edition High Availability Administration Guide.
Determine the absolute number of lost UDP packets after the clusters are started.
If the difference in the number of lost packets is significant, increase the size of the UDP buffer.

To Set the UDP Buffer Size on Linux Systems

On Linux systems, a default UDP buffer size is set for the client, but not for the server. Therefore, on Linux systems, the UDP buffer size might have to be increased. Setting the UDP buffer size involves setting the following kernel parameters:

net.core.rmem_max
net.core.wmem_max
net.core.rmem_default
net.core.wmem_default

Set the kernel parameters in the /etc/sysctl.conf file or at runtime.

If you set the parameters in the /etc/sysctl.conf file, the settings are preserved when the system is rebooted. If you set the parameters at runtime, the settings are not preserved when the system is rebooted.

To set the parameters in the /etc/sysctl.conf file, add or edit the following lines in the file:

net.core.rmem_max=rmem-max
net.core.wmem_max=wmem-max
net.core.rmem_default=rmem-default
net.core.wmem_default=wmem-default

To set the parameters at runtime, use the sysctl command.

$ /sbin/sysctl -w net.core.rmem_max=rmem-max \
net.core.wmem_max=wmem-max \
net.core.rmem_default=rmem-default \
net.core.wmem_default=wmem-default

Example 5-1 Setting the UDP Buffer Size in the /etc/sysctl.conf File

This example shows the lines in the /etc/sysctl.conf file for setting the kernel parameters for controlling the UDP buffer size to 524288.

net.core.rmem_max=524288
net.core.wmem_max=524288
net.core.rmem_default=524288
net.core.wmem_default=524288

Example 5-2 Setting the UDP Buffer Size at Runtime

This example sets the kernel parameters for controlling the UDP buffer size to 524288 at runtime.

$ /sbin/sysctl -w net.core.rmem_max=524288 \
net.core.wmem_max=52428 \
net.core.rmem_default=52428 \
net.core.wmem_default=524288
net.core.rmem_max = 524288
net.core.wmem_max = 52428
net.core.rmem_default = 52428
net.core.wmem_default = 524288

Solaris 10 Platform-Specific Tuning Information

Solaris Dynamic Tracing (DTrace) is a comprehensive dynamic tracing framework for the Solaris Operating System (OS). You can use the DTrace Toolkit to monitor the system. The DTrace Toolkit is available through the OpenSolaris project from the DTraceToolkit page (http://hub.opensolaris.org/bin/view/Community+Group+dtrace/dtracetoolkit).

Tuning Parameters

Tuning Solaris TCP/IP settings benefits programs that open and close many sockets. Since the GlassFish Server operates with a small fixed set of connections, the performance gain might not be significant.

The following table shows Solaris tuning parameters that affect performance and scalability benchmarking. These values are examples of how to tune your system for best performance.

Table 5-1 Tuning Parameters for Solaris

Parameter	Scope	Default	Tuned Value	Comments
`rlim_fd_max`	`/etc/system`	65536	65536	Limit of process open file descriptors. Set to account for expected load (for associated sockets, files, and pipes if any).
`rlim_fd_cur`	`/etc/system`	1024	8192	+
`sq_max_size`	`/etc/system`	2	0	Controls streams driver queue size; setting to 0 makes it infinite so the performance runs won’t be hit by lack of buffer space. Set on clients too. Note that setting `sq_max_size` to 0 might not be optimal for production systems with high network traffic.
`tcp_close_wait_interval`	`ndd /dev/tcp`	240000	60000	Set on clients too.
`tcp_time_wait_interval`	`ndd /dev/tcp`	240000	60000	Set on clients too.
`tcp_conn_req_max_q`	`ndd /dev/tcp`	128	1024	+
`tcp_conn_req_max_q0`	`ndd /dev/tcp`	1024	4096	+
`tcp_ip_abort_interval`	`ndd /dev/tcp`	480000	60000	+
`tcp_keepalive_interval`	`ndd /dev/tcp`	7200000	900000	For high traffic web sites, lower this value.
`tcp_rexmit_interval_initial`	`ndd /dev/tcp`	3000	3000	If retransmission is greater than 30-40%, you should increase this value.
`tcp_rexmit_interval_max`	`ndd /dev/tcp`	240000	10000	+
`tcp_rexmit_interval_min`	`ndd /dev/tcp`	200	3000	+
`tcp_smallest_anon_port`	`ndd /dev/tcp`	32768	1024	Set on clients too.
`tcp_slow_start_initial`	`ndd /dev/tcp`	1	2	Slightly faster transmission of small amounts of data.
`tcp_xmit_hiwat`	`ndd /dev/tcp`	8129	32768	Size of transmit buffer.
`tcp_recv_hiwat`	`ndd /dev/tcp`	8129	32768	Size of receive buffer.
`tcp_conn_hash_size`	`ndd /dev/tcp`	512	8192	Size of connection hash table. See Sizing the Connection Hash Table.

Sizing the Connection Hash Table

The connection hash table keeps all the information for active TCP connections. Use the following command to get the size of the connection hash table:

ndd -get /dev/tcp tcp_conn_hash

This value does not limit the number of connections, but it can cause connection hashing to take longer. The default size is 512.

To make lookups more efficient, set the value to half of the number of concurrent TCP connections that are expected on the server. You can set this value only in /etc/system, and it becomes effective at boot time.

Use the following command to get the current number of TCP connections.

netstat -nP tcp|wc -l

File Descriptor Setting

On the Solaris OS, setting the maximum number of open files property using ulimit has the biggest impact on efforts to support the maximum number of RMI/IIOP clients.

To increase the hard limit, add the following command to /etc/system and reboot it once:

set rlim_fd_max = 8192

Verify this hard limit by using the following command:

ulimit -a -H

Once the above hard limit is set, increase the value of this property explicitly (up to this limit) using the following command:

ulimit -n 8192

Verify this limit by using the following command:

ulimit -a

For example, with the default ulimit of 64, a simple test driver can support only 25 concurrent clients, but with ulimit set to 8192, the same test driver can support 120 concurrent clients. The test driver spawned multiple threads, each of which performed a JNDI lookup and repeatedly called the same business method with a think (delay) time of 500 ms between business method calls, exchanging data of about 100 KB. These settings apply to RMI/IIOP clients on the Solaris OS.

Tuning for Solaris on x86

The following are some options to consider when tuning Solaris on x86 for GlassFish Server:

File Descriptors
IP Stack Settings

Some of the values depend on the system resources available. After making any changes to /etc/system, reboot the machines.

File Descriptors

Add (or edit) the following lines in the /etc/system file:

set rlim_fd_max=65536
set rlim_fd_cur=65536
set sq_max_size=0
set tcp:tcp_conn_hash_size=8192
set autoup=60
set pcisch:pci_stream_buf_enable=0

These settings affect the file descriptors.

IP Stack Settings

Add (or edit) the following lines in the /etc/system file:

set ip:tcp_squeue_wput=1
set ip:tcp_squeue_close=1
set ip:ip_squeue_bind=1
set ip:ip_squeue_worker_wait=10
set ip:ip_squeue_profile=0

These settings tune the IP stack.

To preserve the changes to the file between system reboots, place the following changes to the default TCP variables in a startup script that gets executed when the system reboots:

ndd -set /dev/tcp tcp_time_wait_interval 60000
ndd -set /dev/tcp tcp_conn_req_max_q 16384
ndd -set /dev/tcp tcp_conn_req_max_q0 16384
ndd -set /dev/tcp tcp_ip_abort_interval 60000
ndd -set /dev/tcp tcp_keepalive_interval 7200000
ndd -set /dev/tcp tcp_rexmit_interval_initial 4000
ndd -set /dev/tcp tcp_rexmit_interval_min 3000
ndd -set /dev/tcp tcp_rexmit_interval_max 10000
ndd -set /dev/tcp tcp_smallest_anon_port 32768
ndd -set /dev/tcp tcp_slow_start_initial 2
ndd -set /dev/tcp tcp_xmit_hiwat 32768
ndd -set /dev/tcp tcp_recv_hiwat 32768

Tuning for Linux platforms

To tune for maximum performance on Linux, you need to make adjustments to the following:

Startup Files
File Descriptors
Virtual Memory
Network Interface
Disk I/O Settings
TCP/IP Settings

Startup Files

The following parameters must be added to the /etc/rc.d/rc.local file that gets executed during system startup.

<-- begin
#max file count updated ~256 descriptors per 4Mb.
Specify number of file descriptors based on the amount of system RAM.
echo "6553"> /proc/sys/fs/file-max
#inode-max 3-4 times the file-max
#file not present!!!!!
#echo"262144"> /proc/sys/fs/inode-max
#make more local ports available
echo 1024 25000> /proc/sys/net/ipv4/ip_local_port_range
#increase the memory available with socket buffers
echo 2621143> /proc/sys/net/core/rmem_max
echo 262143> /proc/sys/net/core/rmem_default
#above configuration for 2.4.X kernels
echo 4096 131072 262143> /proc/sys/net/ipv4/tcp_rmem
echo 4096 13107262143> /proc/sys/net/ipv4/tcp_wmem
#disable "RFC2018 TCP Selective Acknowledgements," and
"RFC1323 TCP timestamps" echo 0> /proc/sys/net/ipv4/tcp_sack
echo 0> /proc/sys/net/ipv4/tcp_timestamps
#double maximum amount of memory allocated to shm at runtime
echo "67108864"> /proc/sys/kernel/shmmax
#improve virtual memory VM subsystem of the Linux
echo "100 1200 128 512 15 5000 500 1884 2"> /proc/sys/vm/bdflush
#we also do a sysctl
sysctl -p /etc/sysctl.conf
-- end -->

Additionally, create an /etc/sysctl.conf file and append it with the following values:

<-- begin
 #Disables packet forwarding
net.ipv4.ip_forward = 0
#Enables source route verification
net.ipv4.conf.default.rp_filter = 1
#Disables the magic-sysrq key
kernel.sysrq = 0
fs.file-max=65536
vm.bdflush = 100 1200 128 512 15 5000 500 1884 2
net.ipv4.ip_local_port_range = 1024 65000
net.core.rmem_max= 262143
net.core.rmem_default = 262143
net.ipv4.tcp_rmem = 4096 131072 262143
net.ipv4.tcp_wmem = 4096 131072 262143
net.ipv4.tcp_sack = 0
net.ipv4.tcp_timestamps = 0
kernel.shmmax = 67108864

File Descriptors

You may need to increase the number of file descriptors from the default. Having a higher number of file descriptors ensures that the server can open sockets under high load and not abort requests coming in from clients.

Start by checking system limits for file descriptors with this command:

cat /proc/sys/fs/file-max
8192

The current limit shown is 8192. To increase it to 65535, use the following command (as root):

echo "65535"> /proc/sys/fs/file-max

To make this value to survive a system reboot, add it to /etc/sysctl.conf and specify the maximum number of open files permitted:

fs.file-max = 65535

Note that the parameter is not proc.sys.fs.file-max, as one might expect.

To list the available parameters that can be modified using sysctl:

sysctl -a

To load new values from the sysctl.conf file:

sysctl -p /etc/sysctl.conf

To check and modify limits per shell, use the following command:

limit

The output will look something like this:

cputime         unlimited
filesize        unlimited
datasize        unlimited
stacksize       8192 kbytes
coredumpsize    0 kbytes
memoryuse       unlimited
descriptors     1024
memorylocked    unlimited
maxproc         8146
openfiles       1024

The openfiles and descriptors show a limit of 1024. To increase the limit to 65535 for all users, edit /etc/security/limits.conf as root, and modify or add the nofile setting (number of file) entries:

*         soft    nofile                     65535
*         hard    nofile                     65535

The character "*" is a wildcard that identifies all users. You could also specify a user ID instead.

Then edit /etc/pam.d/login and add the line:

session required /lib/security/pam_limits.so

On Red Hat, you also need to edit /etc/pam.d/sshd and add the following line:

session required /lib/security/pam_limits.so

On many systems, this procedure will be sufficient. Log in as a regular user and try it before doing the remaining steps. The remaining steps might not be required, depending on how pluggable authentication modules (PAM) and secure shell (SSH) are configured.

Virtual Memory

To change virtual memory settings, add the following to /etc/rc.local:

echo 100 1200 128 512 15 5000 500 1884 2> /proc/sys/vm/bdflush

For more information, view the man pages for bdflush.

Network Interface

To ensure that the network interface is operating in full duplex mode, add the following entry into /etc/rc.local:

mii-tool -F 100baseTx-FD eth0

where eth0 is the name of the network interface card (NIC).

Disk I/O Settings

To tune disk I/O performance for non SCSI disks

Test the disk speed.
Use this command:

/sbin/hdparm -t /dev/hdX

Enable direct memory access (DMA).
Use this command:

/sbin/hdparm -d1 /dev/hdX

Check the speed again using the hdparm command.
Given that DMA is not enabled by default, the transfer rate might have improved considerably. In order to do this at every reboot, add the /sbin/hdparm -d1 /dev/hdX line to /etc/conf.d/local.start, /etc/init.d/rc.local, or whatever the startup script is called.
For information on SCSI disks, see: System Tuning for Linux Servers — SCSI (http://people.redhat.com/alikins/system_tuning.html#scsi).

TCP/IP Settings

To tune the TCP/IP settings

Add the following entry to /etc/rc.local

echo 30> /proc/sys/net/ipv4/tcp_fin_timeout
echo 60000> /proc/sys/net/ipv4/tcp_keepalive_time
echo 15000> /proc/sys/net/ipv4/tcp_keepalive_intvl
echo 0> /proc/sys/net/ipv4/tcp_window_scaling

Add the following to /etc/sysctl.conf

# Disables packet forwarding
net.ipv4.ip_forward = 0
# Enables source route verification
net.ipv4.conf.default.rp_filter = 1
# Disables the magic-sysrq key
kernel.sysrq = 0
net.ipv4.ip_local_port_range = 1204 65000
net.core.rmem_max = 262140
net.core.rmem_default = 262140
net.ipv4.tcp_rmem = 4096 131072 262140
net.ipv4.tcp_wmem = 4096 131072 262140
net.ipv4.tcp_sack = 0
net.ipv4.tcp_timestamps = 0
net.ipv4.tcp_window_scaling = 0
net.ipv4.tcp_keepalive_time = 60000
net.ipv4.tcp_keepalive_intvl = 15000
net.ipv4.tcp_fin_timeout = 30

Add the following as the last entry in /etc/rc.local

sysctl -p /etc/sysctl.conf

Reboot the system.
Use this command to increase the size of the transmit buffer:

tcp_recv_hiwat ndd /dev/tcp 8129 32768

Tuning UltraSPARC CMT-Based Systems

Use a combination of tunable parameters and other parameters to tune UltraSPARC CMT-based systems. These values are an example of how you might tune your system to achieve the desired result.

Tuning Operating System and TCP Settings

The following table shows the operating system tuning for Solaris 10 used when benchmarking for performance and scalability on UtraSPARC CMT-based systems (64-bit systems).

Table 5-2 Tuning 64-bit Systems for Performance Benchmarking

Parameter Scope Default Value Tuned Value Comments

Parameter	Scope	Default Value	Tuned Value	Comments
`rlim_fd_max`	`/etc/system`	65536	260000	Process open file descriptors limit; should account for the expected load (for the associated sockets, files, pipes if any).
`hires_tick`	`/etc/system`	+	1	+
`sq_max_size`	`/etc/system`	2	0	Controls streams driver queue size; setting to 0 makes it infinite so the performance runs won’t be hit by lack of buffer space. Set on clients too. Note that setting `sq_max_size` to 0 might not be optimal for production systems with high network traffic.
`ip:ip_squeue_bind`	+	+	0	+
`ip:ip_squeue_fanout`	+	+	1	+
`ipge:ipge_taskq_disable`	`/etc/system`	+	0	+
`ipge:ipge_tx_ring_size`	`/etc/system`	+	2048	+
`ipge:ipge_srv_fifo_depth`	`/etc/system`	+	2048	+
`ipge:ipge_bcopy_thresh`	`/etc/system`	+	384	+
`ipge:ipge_dvma_thresh`	`/etc/system`	+	384	+
`ipge:ipge_tx_syncq`	`/etc/system`	+	1	+
`tcp_conn_req_max_q`	`ndd /dev/tcp`	128	3000	+
`tcp_conn_req_max_q0`	`ndd /dev/tcp`	1024	3000	+
`tcp_max_buf`	`ndd /dev/tcp`	+	4194304	+
`tcp_cwnd_max`	`ndd/dev/tcp`	+	2097152	+
`tcp_xmit_hiwat`	`ndd /dev/tcp`	8129	400000	To increase the transmit buffer.
`tcp_recv_hiwat`	`ndd /dev/tcp`	8129	400000	To increase the receive buffer.

rlim_fd_max

/etc/system

65536

260000

Process open file descriptors limit; should account for the expected load (for the associated sockets, files, pipes if any).

hires_tick

/etc/system

sq_max_size

/etc/system

Controls streams driver queue size; setting to 0 makes it infinite so the performance runs won’t be hit by lack of buffer space. Set on clients too. Note that setting sq_max_size to 0 might not be optimal for production systems with high network traffic.

ip:ip_squeue_bind

ip:ip_squeue_fanout

ipge:ipge_taskq_disable

/etc/system

ipge:ipge_tx_ring_size

/etc/system

2048

ipge:ipge_srv_fifo_depth

/etc/system

2048

ipge:ipge_bcopy_thresh

/etc/system

384

ipge:ipge_dvma_thresh

/etc/system

384

ipge:ipge_tx_syncq

/etc/system

tcp_conn_req_max_q

ndd /dev/tcp

128

3000

tcp_conn_req_max_q0

ndd /dev/tcp

1024

3000

tcp_max_buf

ndd /dev/tcp

4194304

tcp_cwnd_max

ndd/dev/tcp

2097152

tcp_xmit_hiwat

ndd /dev/tcp

8129

400000

To increase the transmit buffer.

tcp_recv_hiwat

ndd /dev/tcp

8129

400000

To increase the receive buffer.

Note that the IPGE driver version is 1.25.25.

Disk Configuration

If HTTP access is logged, follow these guidelines for the disk:

Write access logs on faster disks or attached storage.
If running multiple instances, move the logs for each instance onto separate disks as much as possible.
Enable the disk read/write cache. Note that if you enable write cache on the disk, some writes might be lost if the disk fails.
Consider mounting the disks with the following options, which might yield better disk performance: nologging, directio, noatime.

Network Configuration

If more than one network interface card is used, make sure the network interrupts are not all going to the same core. Run the following script to disable interrupts:

allpsr=`/usr/sbin/psrinfo | grep -v off-line | awk '{ print $1 }'`
   set $allpsr
   numpsr=$#
   while [ $numpsr -gt 0 ];
   do
       shift
       numpsr=`expr $numpsr - 1`
       tmp=1
       while [ $tmp -ne 4 ];
       do
           /usr/sbin/psradm -i $1
           shift
           numpsr=`expr $numpsr - 1`
           tmp=`expr $tmp + 1`
       done
   done

Put all network interfaces into a single group. For example:

$ifconfig ipge0 group webserver
$ifconfig ipge1 group webserver