Fine Tuning/Troubleshooting Ips Iscsi Tcp Performance - Cisco MDS 9000 Series Troubleshooting Manual

Fine Tuning/Troubleshooting IPS iSCSI TCP Performance

S e n d c o m m e n t s t o m d s f e e d b a c k - d o c @ c i s c o . c o m .
Fine Tuning/Troubleshooting IPS iSCSI TCP Performance
Generally there are two segments which will effect the iSCSI performance. First is the FC side flow
control mechanism (Buffer to Buffer Credits, and the FC max frame size) Second is the TCP/IP side.
As in all TCP/IP-related throughput issues, the most important criteria are the Receive/Send Window
Sizes on both TCP end points, RTT (Round Trip Time), actual available bandwidth between the TCP
peers, the MSS (Maximum Segment Size) and the support for higher MTUs between the peers..
The following commands will give you information related to these criteria.
show iscsi remote-node iscsi-session-detail tcp-parameters
show ips stats tcp interface gigabitethernet 2/1 detail
show interface iscsi <x/y>
show interface gigabitethernet <x/y>
show interface fc <x/y>
show iscsi remote-node fcp-session-detail
The default MTU size of an ethernet network is 1500, while the FC networks generally support
maximum frame sizes of 2148 bytes.. This means that an iSCSI gateway will need to chop the FC frames
into two TCP segments or IP fragments while transferring form the FC side to the IP side depending on
how this chopping is implemented within the device.
The IPS module adjusts the Receive Data Field Size that it advertises to its FC partner, according to the
MTU that is configured on the corresponding Gigagbit port of an iSCSI client.
If left to default MTU, the FC frame size from the Target device is decreased to match the maximum
Ethernet frame size, so that the switching of the packet through the switch is swifter. Hence, one point
of performance tuning is increasing the MTU of the IP network between the peers. In this setup there is
one single Catalyst switch.
Jumbo support was enabled for the IPS ports, as well as the MTU for the VLAN corresponding to these
ports was increased.
The second point is to increase the TCP window size of the iSCSI end points. Depending on the latency
between the iSCSI client and IPS, this will need fine tuning. The switch's iSCSI configuration defines
the TCP window size in kilobytes.
Any value starting with 64K ( > 65535 = 0xFFFF bytes) will automatically trigger TCP window scaling
according to RFC1323. The IPS TCP Window scaling begins only when the remote peer (iSCSI client
in this case) requests it. This means that you need to configure the TCP stack of your client to trigger
this functionality (see
For the FC side, depending on the direction of the traffic, the B2Bcredit of the ports corresponding to
the input interfaces (feeding/receiving traffic to/from the iSCSI side) could be increased, especially in
the case of local Gigabit Ethernet attached iSCSI clients.
Each of the above-mentioned commands are taken from a scenario in
sections of the displays are highlighted/italicized or bolded.
Cisco MDS 9000 Family Troubleshooting Guide
5-44
Current retransmission timeout is 300 ms
Round trip time: Smoothed 165 ms, Variance: 35
Advertized window: Current: 61 KB, Maximum: 62 KB, Scale: 0
Peer receive window: Current: 63 KB, Maximum: 63 KB, Scale: 0
Congestion window: Current: 63 KB
Figure 5-14).
Chapter 5 Troubleshooting IP Storage Issues
Figure 5-14. The important
OL-5183-02, Cisco MDS SAN-OS Release 1.3