Configuring Optimization Features : Configuring TCP, satellite optimization, and high-speed TCP
  
Configuring TCP, satellite optimization, and high-speed TCP
This section describes how to configure TCP, satellite optimization, and high-speed TCP settings. It includes these topics:
• Optimizing TCP and satellite WANs
• High-speed TCP optimization
You configure TCP, high-speed TCP, and satellite optimization settings in the Optimization > Network Services: Transport Settings page.
Optimizing TCP and satellite WANs
Riverbed provides satellite WAN optimization to overcome the common sources of performance loss associated with space networking. Satellite optimization allows for more effective use of satellite channels, while providing improved user experiences and increased productivity.
SkipWare, an exclusive technology in the Riverbed product family, senses increases and decreases in bandwidth allocation and automatically adjusts its transmission window in response, without requiring user intervention.
Optimizing SCPS with SkipWare
RiOS includes compatibility settings for the Space Communications Protocol Standards (SCPS) protocol suite. SCPS is designed to allow communication over challenging environments. Originally, it was developed jointly by NASA and DOD’s USSPACECOM to meet their various needs and requirements. Through a collaborative, multiyear R&D effort, the partnership created the Space Communications Protocol Standards-Transport Protocol (SCPS-TP, commonly referred to as “skips”). This protocol now meets the needs of the satellite and wireless communities.
Unlike TCP, the SCPS protocol was designed to operate in an environment of high latency and limited bandwidth. The first commercial implementation of the SCPS protocol was released under the brand name SkipWare.
To use the SkipWare discovery mechanisms, you must have a SCPS license.
• For CX 580, 780, and 3080 appliances, the SCPS license is part of the Enterprise tier.
• For CX 5080 and 7080 appliances, the SCPS license is included with the standard license package (license group 1).
• For x70 and xx70 appliances, the SCPS license is a separate license. You receive a SCPS license when you purchase the SCPS option. If you did not receive your SCPS license, contact your sales representative.
• For Next Generation SteelHead-v (models VCX10-110), the SCPS license is included. For all other SteelHead-v models, the SCPS license is a separate purchase.
SkipWare is enabled automatically when the license is installed, regardless of which transport optimization method is selected (for example, standard TCP, high-speed TCP, or bandwidth estimation). After installing the SkipWare license, you must restart the optimization service.
The basic RiOS license includes non-SkipWare options such as bandwidth estimation and standard TCP.
To change SkipWare settings, you must have role-based permission to use the Optimization Service role. For details, see Managing user permissions.
For details and example satellite deployments, see the SteelHead Deployment Guide.
SCPS connection types
You configure satellite optimization settings depending on the connection type. This section describes the connection types. For details about the SCPS discovery process used in various device scenarios, see the SteelHead Deployment Guide.
RiOS and SCPS connection
A RiOS and SCPS connection is established between two SteelHeads. Because both SteelHeads are SCPS compatible, this is a double-ended connection that benefits from traditional RiOS optimization (SDR and LZ). A RiOS and SCPS connection works with all RiOS features.
RiOS and SCPS Connection
SEI connection
A single-ended interception (SEI) connection is established between a single SteelHead paired with a third-party device running TCP-PEP (Performance Enhancing Proxy). Both the SteelHead and the TCP-PEP device are using the SCPS protocol to speed up the data transfer on a satellite link or other high-latency links. In the following figure, the SteelHead replaces a third-party device running TCP-PEP in the data center, but the SteelHead can also reside in the branch office. Because there’s only one SteelHead that intercepts the connection, this is called a single-ended interception.
Single-Ended Interception Connection
Because a single-ended interception connection communicates with only one SteelHead, it:
• performs only sender-side TCP optimization.
• supports virtual in-path deployments such as WCCP and PBR.
• can’t initiate a SCPS connection on a server-side out-of-path SteelHead.
• supports kickoff.
• supports autodiscovery failover (failover is compatible with IPv6).
• coexists with high-speed TCP.
• doesn’t work with connection forwarding.
Even without a license, you can configure a rule in the SCPS rule table for SEI connections and the SCPS option is added to the SYN packet allowing you to achieve SCPS optimization.
To configure satellite optimization for an SEI, you define SEI connection rules. The SteelHead uses SEI connection rules to determine whether to enable or pass-through SCPS connections.
We recommend that for SEI configurations in which the SteelHead initiates the SCPS connection on the WAN, you add an in-path pass-through rule from the client to the server. While the pass-through rule is optional, without it the SteelHead probes for another SteelHead, and when it doesn’t locate one, will failover. Adding the in-path pass-through rule speeds up setup by eliminating the autodiscovery probe and subsequent failover.
The in-path pass-through rule isn’t necessary on SEI configurations in which the SteelHead terminates the SCPS connection on the WAN, because in this configuration the SteelHead evaluates only the SEI connection rules table and ignores the in-path rules table.
SEI connections count toward the connection count limit on the SteelHead.
When server-side network asymmetry occurs in an SEI configuration, the server-side SteelHead creates a bad RST log entry in the asymmetric routing table. This log entry differs from other configurations (non-SCPS) in that the client-side SteelHead typically detects asymmetry because of the bad RST and creates an entry in the asymmetric routing table. In SEI configurations, the SteelHead detects asymmetry and creates asymmetric routing table entries independent of other SteelHeads. This results in a TCP proxy only connection between the client-side SteelHead and the server when autodiscovery is disabled. For details about the asymmetric routing table, see Configuring asymmetric routing features.
To configure TCP and SkipWare SCPS optimization
To properly configure transport settings for your environment, you must understand its characteristics. For information on gathering performance characteristics for your environment, see the SteelHead Deployment Guide.
1. Choose Optimization > Network Services: Transport Settings to display the Transport Settings page.
Transport Settings page
2. Under TCP Optimization, complete the configuration as described in this table.
Control
Description
Auto-Detect
Automatically detects the optimal TCP configuration by using the same mode as the peer SteelHead for inner connections, SkipWare when negotiated, or standard TCP for all other cases. This is the default setting.
If you have a mixed environment where several different types of networks terminate into a hub or server-side SteelHead, enable this setting on your hub SteelHead so it can reflect the various transport optimization mechanisms of your remote site SteelHeads. Otherwise, you can hard code your hub SteelHead to the desired setting.
RiOS advertises automatic detection of TCP optimization to a peer SteelHead through the OOB connection between the appliances.
For single-ended interception connections, use SkipWare per-connection TCP optimization when possible; use standard TCP otherwise.
Standard (RFC-Compliant)
Optimizes non-SCPS TCP connections by applying data and transport streamlining for TCP traffic over the WAN. This control forces peers to use standard TCP as well. For details on data and transport streamlining, see the SteelHead Deployment Guide. This option clears any advanced bandwidth congestion control that was previously set.
HighSpeed
Enables high-speed TCP optimization for more complete use of long fat pipes (high-bandwidth, high-delay networks). Do not enable for satellite networks.
We recommend that you enable high-speed TCP optimization only after you have carefully evaluated whether it will benefit your network environment. For details about the trade-offs of enabling high-speed TCP, see tcp highspeed enable in the Riverbed Command-Line Interface Reference Manual.
Bandwidth Estimation
Uses an intelligent bandwidth estimation algorithm along with a modified slow-start algorithm to optimize performance in long lossy networks. These networks typically include satellite and other wireless environments, such as cellular networks, longer microwave, or Wi-Max networks.
Bandwidth estimation is a sender-side modification of TCP and is compatible with the other TCP stacks in the RiOS system. The intelligent bandwidth estimation is based on analysis of both ACKs and latency measurements. The modified slow-start mechanism enables a flow to ramp up faster in high-latency environments than traditional TCP. The intelligent bandwidth estimation algorithm allows it to learn effective rates for use during modified slow start, and also to differentiate BER loss from congestion-derived loss and manage them accordingly. Bandwidth estimation has good fairness and friendliness qualities toward other traffic along the path.
The default setting is off.
SkipWare Per-Connection
Applies TCP congestion control to each SCPS-capable connection. The congestion control uses:
• a pipe algorithm that gates when a packet should be sent after receipt of an ACK.
• the NewReno algorithm, which includes the sender's congestion window, slow start, and congestion avoidance.
• time stamps, window scaling, appropriate byte counting, and loss detection.
This transport setting uses a modified slow-start algorithm and a modified congestion-avoidance approach. This method enables SCPS per-connection to ramp up flows faster in high-latency environments, and handle lossy scenarios, while remaining reasonably fair and friendly to other traffic. SCPS per-connection does a very good job of efficiently filling up satellite links of all sizes. SCPS per-connection is a high-performance option for satellite networks.
We recommend enabling per-connection if the error rate in the link is less than approximately 1 percent.
The Management Console dims this setting until you install a SkipWare license.
SkipWare Error-Tolerant
Enables SkipWare optimization with the error-rate detection and recovery mechanism on the SteelHead.
This setting allows the per-connection congestion control to tolerate some loss due to corrupted packets (bit errors), without reducing the throughput, using a modified slow-start algorithm and a modified congestion-avoidance approach. It requires significantly more retransmitted packets to trigger this congestion-avoidance algorithm than the SkipWare per-connection setting.
Error-tolerant TCP optimization assumes that the environment has a high BER and that most retransmissions are due to poor signal quality instead of congestion. This method maximizes performance in high-loss environments, without incurring the additional per-packet overhead of a FEC algorithm at the transport layer.
SCPS error tolerance is a high-performance option for lossy satellite networks.
Use caution when enabling error-tolerant TCP optimization, particularly in channels with coexisting TCP traffic, because it can be quite aggressive and adversely affect channel congestion with competing TCP flows.
We recommend enabling error tolerance if the error rate in the link is more than approximately 1 percent.
The Management Console dims this setting until you install a SkipWare license.
Cubic
Enables the Cubic congestion control algorithm. Cubic is the local default congestion control algorithm when two peer SteelHeads are both configured to auto-detect. Cubic offers better performance and faster recovery after congestion events than NewReno, the previous local default.
Enable Rate Pacing
Imposes a global data-transmit limit on the link rate for all SCPS connections between peer SteelHeads, or on the link rate for a SteelHead paired with a third-party device running TCP-PEP (Performance Enhancing Proxy).
Rate pacing combines MX-TCP and a congestion-control method of your choice for connections between peer SteelHeads and SEI connections (on a per-rule basis). The congestion-control method runs as an overlay on top of MX-TCP and probes for the actual link rate. It then communicates the available bandwidth to MX-TCP.
Enable rate pacing to prevent these problems:
• Congestion loss while exiting the slow-start phase. The slow-start phase is an important part of the TCP congestion-control mechanisms that starts slowly increasing its window size as it gains confidence about the network throughput.
• Congestion collapse
• Packet bursts
Rate pacing is disabled by default.
With no congestion, the slow-start phase ramps up to the MX-TCP rate and settles there. When RiOS detects congestion (either due to other sources of traffic, a bottleneck other than the satellite modem, or because of a variable modem rate), the congestion-control method kicks in to avoid congestion loss and exit the slow-start phase faster.
Enable rate pacing on the client-side SteelHead along with a congestion-control method. The client-side SteelHead communicates to the server-side SteelHead that rate pacing is in effect. You must also:
• Enable Auto-Detect TCP Optimization on the server-side SteelHead to negotiate the configuration with the client-side SteelHead.
• Configure an MX-TCP QoS rule to set the appropriate rate cap. If an MX-TCP QoS rule is not in place, the system doesn’t apply rate pacing and the congestion-control method takes effect. You can’t delete the MX-TCP QoS rule when rate pacing is enabled.
The Management Console dims this feature until you install a SkipWare license.
Rate pacing doesn’t support IPv6.
You can also enable rate pacing for SEI connections by defining an SEI rule for each connection.
Enable Single-Ended Connection Rules Table
Enables transport optimization for single-ended interception connections with no SteelHead peer. These connections appear in the rules table.
In RiOS 8.5 or later, you can impose rate pacing for single-ended interception connections with no peer SteelHead. By defining an SEI connection rule, you can enforce rate pacing even when the SteelHead is not peered with a SCPS device and SCPS is not negotiated.
To enforce rate pacing for a single-ended interception connection, create an SEI connection rule for use as a transport-optimization proxy, select a congestion method for the rule, and then configure a QoS rule (with the same client/server subnet) to use MX-TCP. RiOS 8.5 and later accelerate the WAN-originated or LAN-originated proxied connection using MX-TCP.
By default, the SEI connection rules table is disabled. When enabled, two default rules appear in the rules table. The first default rule matches all traffic with the destination port set to the interactive port label and bypasses the connection for SCPS optimization.
The second default rule matches all traffic with the destination port set to the RBT-Proto port label and bypasses the connection for SCPS optimization.
This option doesn’t affect the optimization of SCPS connections between SteelHeads.
When you disable the table, you can still add, move, or remove rules, but the changes don’t take effect until you reenable the table.
The Management Console dims the SEI rules table until you install a SkipWare license.
Enable SkipWare Legacy Compression—Enables negotiation of SCPS-TP TCP header and data compression with a remote SCPS-TP device. This feature enables interoperation with RSP SkipWare packages and TurboIP devices that have also been configured to negotiate TCP header and data compression.
Legacy compression is disabled by default.
After enabling or disabling legacy compression, you must restart the optimization service.
The Management Console dims legacy compression until you install a SkipWare license and enable the SEI rules table.
Legacy compression also works with non-SCPS TCP algorithms.
These limits apply to legacy compression:
• This feature is not compatible with IPv6.
• Packets with a compressed TCP header use IP protocol 105 in the encapsulating IP header; this might require changes to intervening firewalls to permit protocol 105 packets to pass.
• This feature supports a maximum of 255 connections between any pair of end-host IP addresses. The connection limit for legacy SkipWare connections is the same as the appliance-connection limit.
• QoS limits for the SteelHead apply to the legacy SkipWare connections.
To view SCPS connections, see Viewing Connection History reports.
3. Click Apply to save your settings to the running configuration.
4. Click Save to Disk to save your settings permanently.
5. Click Restart Services to restart the optimization service.
Configuring buffer settings
The buffer settings in the Transport Settings page support high-speed TCP and are also used in data protection scenarios to improve performance. For details about data protection deployments, see the SteelHead Deployment Guide.
To properly configure buffer settings for a satellite environment, you must understand its characteristics. For information on gathering performance characteristics for your environment, see the SteelHead Deployment Guide.
To configure buffer settings
1. Choose Optimization > Network Services: Transport Settings to display the Transport Settings page.
2. Under Buffer Settings, complete the configuration as described in this table.
Control
Description
LAN Send Buffer Size
Specify the send buffer size used to send data out of the LAN. The default value is 81920.
LAN Receive Buffer Size
Specify the receive buffer size used to receive data from the LAN. The default value is 32768.
WAN Default Send Buffer Size
Specify the send buffer size used to send data out of the WAN. The default value is 262140.
WAN Default Receive Buffer Size
Specify the receive buffer size used to receive data from the WAN. The default value is 262140.
3. Click Apply to save your settings to the running configuration.
4. Click Save to Disk to save your settings permanently.
Adding single-ended connection rules
You can optionally add rules to control single-ended SCPS connections. The SteelHead uses these rules to determine whether to enable or pass through SCPS connections.
A SteelHead receiving a SCPS connection on the WAN evaluates only the single-ended connection rules table.
To pass through a SCPS connection, we recommend setting both an in-path rule and a single-ended connection rule.
To add a single-ended connection rule
1. Choose Optimization > Network Services: Transport Settings to display the Transport Settings page.
Single-Ended Connection Rules
2. Under Single-Ended Connection Rules, complete the configuration as described in this table.
Control
Description
Add New Rule
Displays the controls for adding a new rule.
Position
Select Start, End, or a rule number from the drop-down list. SteelHeads evaluate rules in numerical order starting with rule 1. If the conditions set in the rule match, then the rule is applied, and the system moves on to the next packet. If the conditions set in the rule don’t match, the system consults the next rule. As an example, if the conditions of rule 1 don’t match, rule 2 is consulted. If rule 2 matches the conditions, it’s applied, and no further rules are consulted.
Source Subnet
Specify an IPv4 or IPv6 address and mask for the traffic source.
You can also specify wildcards:
• All-IPv4 is the wildcard for single-stack IPv4 networks.
• All-IPv6 is the wildcard for single-stack IPv6 networks.
• All-IP is the wildcard for all IPv4 and IPv6 networks.
Use these formats:
xxx.xxx.xxx.xxx/xx (IPv4)
x:x:x::x/xxxx (IPv6)
Destination Subnet
Specify an IPv4 or IPv6 address and mask for the traffic destination.
You can also specify wildcards:
• All-IPv4 is the wildcard for single-stack IPv4 networks.
• All-IPv6 is the wildcard for single-stack IPv6 networks.
• All-IP is the wildcard for all IPv4 and IPv6 networks.
Use these formats:
xxx.xxx.xxx.xxx/xx (IPv4)
x:x:x::x/xxxx (IPv6)
Port or Port Label
Specify the destination port number, port label, or all.
Click Port Label to go to the Networking > App Definitions: Port Labels page for reference.
VLAN Tag ID
Specify one of the following: a VLAN identification number from 1 to 4094; all to specify that the rule applies to all VLANs; or untagged to specify the rule applies to untagged connections.
RiOS supports VLAN v802.1Q. To configure VLAN tagging, configure SCPS rules to apply to all VLANs or to a specific VLAN. By default, rules apply to all VLAN values unless you specify a particular VLAN ID. Pass-through traffic maintains any preexisting VLAN tagging between the LAN and WAN interfaces.
Traffic
Specifies the action that the rule takes on a SCPS connection. To allow single-ended interception SCPS connections to pass through the SteelHead unoptimized, disable SCPS Discover and TCP Proxy.
Select one of these options:
– SCPS Discover—Enables SCPS and disables TCP proxy.
– TCP Proxy—Disables SCPS and enables TCP proxy.
Congestion Control Algorithm
Select a method for congestion control from the drop-down list.
• Standard (RFC-Compliant)—Optimizes non-SCPS TCP connections by applying data and transport streamlining for TCP traffic over the WAN. This control forces peers to use standard TCP as well. For details on data and transport streamlining, see the SteelHead Deployment Guide. This option clears any advanced bandwidth congestion control that was previously set.
• HighSpeed—Enables high-speed TCP optimization for more complete use of long fat pipes (high-bandwidth, high-delay networks). Do not enable for satellite networks.
We recommend that you enable high-speed TCP optimization only after you have carefully evaluated whether it will benefit your network environment. For details about the trade-offs of enabling high-speed TCP, see tcp highspeed enable in the Riverbed Command-Line Interface Reference Manual.
• Bandwidth Estimation—Uses an intelligent bandwidth estimation algorithm along with a modified slow-start algorithm to optimize performance in long lossy networks. These networks typically include satellite and other wireless environments, such as cellular networks, longer microwave, or Wi-Max networks.
Bandwidth estimation is a sender-side modification of TCP and is compatible with the other TCP stacks in the RiOS system. The intelligent bandwidth estimation is based on analysis of both ACKs and latency measurements. The modified slow-start mechanism enables a flow to ramp up faster in high latency environments than traditional TCP. The intelligent bandwidth estimation algorithm allows it to learn effective rates for use during modified slow start, and also to differentiate BER loss from congestion-derived loss and deal with them accordingly. Bandwidth estimation has good fairness and friendliness qualities toward other traffic along the path.
• SkipWare Per-Connection—Applies TCP congestion control to each SCPS-capable connection. This method is compatible with IPv6. The congestion control uses:
– a pipe algorithm that gates when a packet should be sent after receipt of an ACK.
– the NewReno algorithm, which includes the sender's congestion window, slow start, and congestion avoidance.
– time stamps, window scaling, appropriate byte counting, and loss detection.
This transport setting uses a modified slow-start algorithm and a modified congestion-avoidance approach. This method enables SCPS per connection to ramp up flows faster in high-latency environments, and handle lossy scenarios, while remaining reasonably fair and friendly to other traffic. SCPS per-connection does a very good job of efficiently filling up satellite links of all sizes. SkipWare per-connection is a high-performance option for satellite networks.
The Management Console dims this setting until you install a SkipWare license.
 
• SkipWare Error-Tolerant—Enables SkipWare optimization with the error-rate detection and recovery mechanism on the SteelHead. This method is compatible with IPv6.
This method tolerates some loss due to corrupted packets (bit errors), without reducing the throughput, using a modified slow-start algorithm and a modified congestion avoidance approach. It requires significantly more retransmitted packets to trigger this congestion-avoidance algorithm than the SkipWare per-connection setting. Error-tolerant TCP optimization assumes that the environment has a high BER and most retransmissions are due to poor signal quality instead of congestion. This method maximizes performance in high-loss environments, without incurring the additional per-packet overhead of a FEC algorithm at the transport layer.
Use caution when enabling error-tolerant TCP optimization, particularly in channels with coexisting TCP traffic, because it can be quite aggressive and adversely affect channel congestion with competing TCP flows.
The Management Console dims this setting until you install a SkipWare license.
Enable Rate Pacing
Imposes a global data transmit limit on the link rate for all SCPS connections between peer SteelHeads or on the link rate for a SteelHead paired with a third-party device running TCP-PEP (Performance Enhancing Proxy).
Rate pacing combines MX-TCP and a congestion-control method of your choice for connections between peer SteelHeads and SEI connections (on a per-rule basis). The congestion-control method runs as an overlay on top of MX-TCP and probes for the actual link rate. It then communicates the available bandwidth to MX-TCP.
Enable rate pacing to prevent these problems:
• Congestion loss while exiting the slow start phase. The slow-start phase is an important part of the TCP congestion-control mechanisms that starts slowly increasing its window size as it gains confidence about the network throughput.
• Congestion collapse.
• Packet bursts.
Rate pacing is disabled by default.
With no congestion, the slow start ramps up to the MX-TCP rate and settles there. When RiOS detects congestion (either due to other sources of traffic, a bottleneck other than the satellite modem, or because of a variable modem rate), the congestion-control method kicks in to avoid congestion loss and exit the slow start phase faster.
Enable rate pacing on the client-side SteelHead along with a congestion-control method. The client-side SteelHead communicates to the server-side SteelHead that rate pacing is in effect. You must also:
• Enable Auto-Detect TCP Optimization on the server-side SteelHead to negotiate the configuration with the client-side SteelHead.
• Configure an MX-TCP QoS rule to set the appropriate rate cap. If an MX-TCP QoS rule is not in place, rate pacing is not applied and the congestion-control method takes effect. You can’t delete the MX-TCP QoS rule when rate pacing is enabled.
The Management Console dims this setting until you install a SkipWare license.
Rate pacing doesn’t support IPv6.
You can also enable rate pacing for SEI connections by defining an SEI rule for each connection.
Add
Adds the rule to the list. The Management Console redisplays the SCPS Rules table and applies your modifications to the running configuration, which is stored in memory.
Remove Selected Rules
Select the check box next to the name and click Remove Selected.
Move Selected Rules
Moves the selected rules. Click the arrow next to the desired rule position; the rule moves to the new position.
3. Click Apply to save your settings to the running configuration.
4. Click Save to Disk to save your settings permanently.
After you apply your settings, you can verify whether changes have had the desired effect by viewing the Current Connections report. The report summarizes the optimized established connections for SCPS. SCPS connections appear as typical established, optimized or established, or single-ended optimized connections. Click the connection to view details. SCPS connection detail reports display SCPS Initiate or SCPS Terminate under Connection Information. Under Congestion Control, the report displays the congestion control method that the connection is using.
High-speed TCP optimization
The high-speed TCP (HS-TCP) feature provides acceleration and high throughput for high-bandwidth links (also known as long fat networks, or LFNs) for which the WAN pipe is large but latency is high. HS-TCP is activated for all connections that have a BDP larger than 100 packets.
For details about using HS-TCP in data protection scenarios, see the SteelHead Deployment Guide.
HS-TCP basic steps
This table describes the basic steps needed to configure high-speed TCP.
Task
Reference
1. Enable high-speed TCP support.
2. Increase the WAN buffers to 2 * Bandwidth Delay Product (BDP).
You can calculate the BDP WAN buffer size:
Buffer size in bytes = 2 * bandwidth (in bits per sec) * delay (in sec) / 8 (bits per byte)
Example: For a link of 155 Mbps and 100 ms round-trip delay.
Bandwidth = 155 Mbps = 155000000 bps
Delay = 100 ms = 0.1 sec
BDP = 155 000 000 * 0.1 / 8 = 1937500 bytes Buffer size in bytes = 2 * BDP = 2 * 1937500 = 3 875 000 bytes.
If this number is greater than the default (256 KB), enable HS-TCP with the correct buffer size.
 
3. Increase the LAN buffers to 1 MB.
4. Enable in-path support.