view Documentation/networking/cxgb.txt @ 897:329ea0ccb344

balloon: try harder to balloon up under memory pressure.

Currently if the balloon driver is unable to increase the guest's
reservation it assumes the failure was due to reaching its full
allocation, gives up on the ballooning operation and records the limit
it reached as the "hard limit". The driver will not try again until
the target is set again (even to the same value).

However it is possible that ballooning has in fact failed due to
memory pressure in the host and therefore it is desirable to keep
attempting to reach the target in case memory becomes available. The
most likely scenario is that some guests are ballooning down while
others are ballooning up and therefore there is temporary memory
pressure while things stabilise. You would not expect a well behaved
toolstack to ask a domain to balloon to more than its allocation nor
would you expect it to deliberately over-commit memory by setting
balloon targets which exceed the total host memory.

This patch drops the concept of a hard limit and causes the balloon
driver to retry increasing the reservation on a timer in the same
manner as when decreasing the reservation.

Also if we partially succeed in increasing the reservation
(i.e. receive less pages than we asked for) then we may as well keep
those pages rather than returning them to Xen.

Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
author Keir Fraser <keir.fraser@citrix.com>
date Fri Jun 05 14:01:20 2009 +0100 (2009-06-05)
parents 831230e53067
line source
1 Chelsio N210 10Gb Ethernet Network Controller
3 Driver Release Notes for Linux
5 Version 2.1.1
7 June 20, 2005
10 ========
20 ============
22 This document describes the Linux driver for Chelsio 10Gb Ethernet Network
23 Controller. This driver supports the Chelsio N210 NIC and is backward
24 compatible with the Chelsio N110 model 10Gb NICs.
28 ========
30 Adaptive Interrupts (adaptive-rx)
31 ---------------------------------
33 This feature provides an adaptive algorithm that adjusts the interrupt
34 coalescing parameters, allowing the driver to dynamically adapt the latency
35 settings to achieve the highest performance during various types of network
36 load.
38 The interface used to control this feature is ethtool. Please see the
39 ethtool manpage for additional usage information.
41 By default, adaptive-rx is disabled.
42 To enable adaptive-rx:
44 ethtool -C <interface> adaptive-rx on
46 To disable adaptive-rx, use ethtool:
48 ethtool -C <interface> adaptive-rx off
50 After disabling adaptive-rx, the timer latency value will be set to 50us.
51 You may set the timer latency after disabling adaptive-rx:
53 ethtool -C <interface> rx-usecs <microseconds>
55 An example to set the timer latency value to 100us on eth0:
57 ethtool -C eth0 rx-usecs 100
59 You may also provide a timer latency value while disabling adpative-rx:
61 ethtool -C <interface> adaptive-rx off rx-usecs <microseconds>
63 If adaptive-rx is disabled and a timer latency value is specified, the timer
64 will be set to the specified value until changed by the user or until
65 adaptive-rx is enabled.
67 To view the status of the adaptive-rx and timer latency values:
69 ethtool -c <interface>
72 TCP Segmentation Offloading (TSO) Support
73 -----------------------------------------
75 This feature, also known as "large send", enables a system's protocol stack
76 to offload portions of outbound TCP processing to a network interface card
77 thereby reducing system CPU utilization and enhancing performance.
79 The interface used to control this feature is ethtool version 1.8 or higher.
80 Please see the ethtool manpage for additional usage information.
82 By default, TSO is enabled.
83 To disable TSO:
85 ethtool -K <interface> tso off
87 To enable TSO:
89 ethtool -K <interface> tso on
91 To view the status of TSO:
93 ethtool -k <interface>
97 ===========
99 The following information is provided as an example of how to change system
100 parameters for "performance tuning" an what value to use. You may or may not
101 want to change these system parameters, depending on your server/workstation
102 application. Doing so is not warranted in any way by Chelsio Communications,
103 and is done at "YOUR OWN RISK". Chelsio will not be held responsible for loss
104 of data or damage to equipment.
106 Your distribution may have a different way of doing things, or you may prefer
107 a different method. These commands are shown only to provide an example of
108 what to do and are by no means definitive.
110 Making any of the following system changes will only last until you reboot
111 your system. You may want to write a script that runs at boot-up which
112 includes the optimal settings for your system.
114 Setting PCI Latency Timer:
115 setpci -d 1425:* 0x0c.l=0x0000F800
117 Disabling TCP timestamp:
118 sysctl -w net.ipv4.tcp_timestamps=0
120 Disabling SACK:
121 sysctl -w net.ipv4.tcp_sack=0
123 Setting large number of incoming connection requests:
124 sysctl -w net.ipv4.tcp_max_syn_backlog=3000
126 Setting maximum receive socket buffer size:
127 sysctl -w net.core.rmem_max=1024000
129 Setting maximum send socket buffer size:
130 sysctl -w net.core.wmem_max=1024000
132 Set smp_affinity (on a multiprocessor system) to a single CPU:
133 echo 1 > /proc/irq/<interrupt_number>/smp_affinity
135 Setting default receive socket buffer size:
136 sysctl -w net.core.rmem_default=524287
138 Setting default send socket buffer size:
139 sysctl -w net.core.wmem_default=524287
141 Setting maximum option memory buffers:
142 sysctl -w net.core.optmem_max=524287
144 Setting maximum backlog (# of unprocessed packets before kernel drops):
145 sysctl -w net.core.netdev_max_backlog=300000
147 Setting TCP read buffers (min/default/max):
148 sysctl -w net.ipv4.tcp_rmem="10000000 10000000 10000000"
150 Setting TCP write buffers (min/pressure/max):
151 sysctl -w net.ipv4.tcp_wmem="10000000 10000000 10000000"
153 Setting TCP buffer space (min/pressure/max):
154 sysctl -w net.ipv4.tcp_mem="10000000 10000000 10000000"
156 TCP window size for single connections:
157 The receive buffer (RX_WINDOW) size must be at least as large as the
158 Bandwidth-Delay Product of the communication link between the sender and
159 receiver. Due to the variations of RTT, you may want to increase the buffer
160 size up to 2 times the Bandwidth-Delay Product. Reference page 289 of
161 "TCP/IP Illustrated, Volume 1, The Protocols" by W. Richard Stevens.
162 At 10Gb speeds, use the following formula:
163 RX_WINDOW >= 1.25MBytes * RTT(in milliseconds)
164 Example for RTT with 100us: RX_WINDOW = (1,250,000 * 0.1) = 125,000
165 RX_WINDOW sizes of 256KB - 512KB should be sufficient.
166 Setting the min, max, and default receive buffer (RX_WINDOW) size:
167 sysctl -w net.ipv4.tcp_rmem="<min> <default> <max>"
169 TCP window size for multiple connections:
170 The receive buffer (RX_WINDOW) size may be calculated the same as single
171 connections, but should be divided by the number of connections. The
172 smaller window prevents congestion and facilitates better pacing,
173 especially if/when MAC level flow control does not work well or when it is
174 not supported on the machine. Experimentation may be necessary to attain
175 the correct value. This method is provided as a starting point fot the
176 correct receive buffer size.
177 Setting the min, max, and default receive buffer (RX_WINDOW) size is
178 performed in the same manner as single connection.
182 ===============
184 The following messages are the most common messages logged by syslog. These
185 may be found in /var/log/messages.
187 Driver up:
188 Chelsio Network Driver - version 2.1.1
190 NIC detected:
191 eth#: Chelsio N210 1x10GBaseX NIC (rev #), PCIX 133MHz/64-bit
193 Link up:
194 eth#: link is up at 10 Gbps, full duplex
196 Link down:
197 eth#: link is down
201 ============
203 These issues have been identified during testing. The following information
204 is provided as a workaround to the problem. In some cases, this problem is
205 inherent to Linux or to a particular Linux Distribution and/or hardware
206 platform.
208 1. Large number of TCP retransmits on a multiprocessor (SMP) system.
210 On a system with multiple CPUs, the interrupt (IRQ) for the network
211 controller may be bound to more than one CPU. This will cause TCP
212 retransmits if the packet data were to be split across different CPUs
213 and re-assembled in a different order than expected.
215 To eliminate the TCP retransmits, set smp_affinity on the particular
216 interrupt to a single CPU. You can locate the interrupt (IRQ) used on
217 the N110/N210 by using ifconfig:
218 ifconfig <dev_name> | grep Interrupt
219 Set the smp_affinity to a single CPU:
220 echo 1 > /proc/irq/<interrupt_number>/smp_affinity
222 It is highly suggested that you do not run the irqbalance daemon on your
223 system, as this will change any smp_affinity setting you have applied.
224 The irqbalance daemon runs on a 10 second interval and binds interrupts
225 to the least loaded CPU determined by the daemon. To disable this daemon:
226 chkconfig --level 2345 irqbalance off
228 By default, some Linux distributions enable the kernel feature,
229 irqbalance, which performs the same function as the daemon. To disable
230 this feature, add the following line to your bootloader:
231 noirqbalance
233 Example using the Grub bootloader:
234 title Red Hat Enterprise Linux AS (2.4.21-27.ELsmp)
235 root (hd0,0)
236 kernel /vmlinuz-2.4.21-27.ELsmp ro root=/dev/hda3 noirqbalance
237 initrd /initrd-2.4.21-27.ELsmp.img
239 2. After running insmod, the driver is loaded and the incorrect network
240 interface is brought up without running ifup.
242 When using 2.4.x kernels, including RHEL kernels, the Linux kernel
243 invokes a script named "hotplug". This script is primarily used to
244 automatically bring up USB devices when they are plugged in, however,
245 the script also attempts to automatically bring up a network interface
246 after loading the kernel module. The hotplug script does this by scanning
247 the ifcfg-eth# config files in /etc/sysconfig/network-scripts, looking
248 for HWADDR=<mac_address>.
250 If the hotplug script does not find the HWADDRR within any of the
251 ifcfg-eth# files, it will bring up the device with the next available
252 interface name. If this interface is already configured for a different
253 network card, your new interface will have incorrect IP address and
254 network settings.
256 To solve this issue, you can add the HWADDR=<mac_address> key to the
257 interface config file of your network controller.
259 To disable this "hotplug" feature, you may add the driver (module name)
260 to the "blacklist" file located in /etc/hotplug. It has been noted that
261 this does not work for network devices because the net.agent script
262 does not use the blacklist file. Simply remove, or rename, the net.agent
263 script located in /etc/hotplug to disable this feature.
265 3. Transport Protocol (TP) hangs when running heavy multi-connection traffic
266 on an AMD Opteron system with HyperTransport PCI-X Tunnel chipset.
268 If your AMD Opteron system uses the AMD-8131 HyperTransport PCI-X Tunnel
269 chipset, you may experience the "133-Mhz Mode Split Completion Data
270 Corruption" bug identified by AMD while using a 133Mhz PCI-X card on the
271 bus PCI-X bus.
273 AMD states, "Under highly specific conditions, the AMD-8131 PCI-X Tunnel
274 can provide stale data via split completion cycles to a PCI-X card that
275 is operating at 133 Mhz", causing data corruption.
277 AMD's provides three workarounds for this problem, however, Chelsio
278 recommends the first option for best performance with this bug:
280 For 133Mhz secondary bus operation, limit the transaction length and
281 the number of outstanding transactions, via BIOS configuration
282 programming of the PCI-X card, to the following:
284 Data Length (bytes): 1k
285 Total allowed outstanding transactions: 2
287 Please refer to AMD 8131-HT/PCI-X Errata 26310 Rev 3.08 August 2004,
288 section 56, "133-MHz Mode Split Completion Data Corruption" for more
289 details with this bug and workarounds suggested by AMD.
291 It may be possible to work outside AMD's recommended PCI-X settings, try
292 increasing the Data Length to 2k bytes for increased performance. If you
293 have issues with these settings, please revert to the "safe" settings
294 and duplicate the problem before submitting a bug or asking for support.
296 NOTE: The default setting on most systems is 8 outstanding transactions
297 and 2k bytes data length.
299 4. On multiprocessor systems, it has been noted that an application which
300 is handling 10Gb networking can switch between CPUs causing degraded
301 and/or unstable performance.
303 If running on an SMP system and taking performance measurements, it
304 is suggested you either run the latest netperf-2.4.0+ or use a binding
305 tool such as Tim Hockin's procstate utilities (runon)
306 <http://www.hockin.org/~thockin/procstate/>.
308 Binding netserver and netperf (or other applications) to particular
309 CPUs will have a significant difference in performance measurements.
310 You may need to experiment which CPU to bind the application to in
311 order to achieve the best performance for your system.
313 If you are developing an application designed for 10Gb networking,
314 please keep in mind you may want to look at kernel functions
315 sched_setaffinity & sched_getaffinity to bind your application.
317 If you are just running user-space applications such as ftp, telnet,
318 etc., you may want to try the runon tool provided by Tim Hockin's
319 procstate utility. You could also try binding the interface to a
320 particular CPU: runon 0 ifup eth0
324 =======
326 If you have problems with the software or hardware, please contact our
327 customer support team via email at support@chelsio.com or check our website
328 at http://www.chelsio.com
330 ===============================================================================
332 Chelsio Communications
333 370 San Aleso Ave.
334 Suite 100
335 Sunnyvale, CA 94085
336 http://www.chelsio.com
338 This program is free software; you can redistribute it and/or modify
339 it under the terms of the GNU General Public License, version 2, as
340 published by the Free Software Foundation.
342 You should have received a copy of the GNU General Public License along
343 with this program; if not, write to the Free Software Foundation, Inc.,
344 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
350 Copyright (c) 2003-2005 Chelsio Communications. All rights reserved.
352 ===============================================================================