HP ATM 622 and 155 MBPS adapters for HP 9000 Enterprise Servers and Workstations

*One CPU, Classical IP Interface, MTU 9180 bytes

*N4000 system, One CPU, 155 Mbps Adapter

*N4000 system, 1 CPU, 622 Mbps Adapter

The TCP window size is an important parameter that affects throughput performance. For 622 Mbps adapters, the TCP window size on the receiving end must be set to at least 144Kbytes (instead of the default window size of 64 Kbytes) to achieve the maximum throughput. For 155 and 622 Mbps adapters, the window size of 144 Kbytes also minimizes the sensitivity to application messages size. See Figure 1 below for an example of Outbound throughput.

Figure 1: TCP Window Size and Outbound TCP throughput on N4000 (Classical IP, 622 Adapter)

The second set of tests consists of exchange of a single request and response packet of 1 byte each. The packet is at the TCP level. The performance metric is the aggregate number of request/response packet pairs (transactions) per second. This performance metric is an indication of the number of user level packets the adapter can process in a second.

Test results show that the CPU and the platform type influence the single byte request/response performance.

A transaction consisting of a request size of 1024 bytes and a response size of 2048 bytes over a UDP protocol models an environment of NFS services running over ATM.

*Classical IP, One CPU
¹ The V2500 Platform tests for 155 and 622 adapters were run with 2 CPUs because the minimum configuration for V2500 is 2 CPUs. The CPU utilization reported for V2500 is per CPU.

*Classical IP, One CPU

The Bidirectional TCP throughput on the V2250 is limited to 680 Mbps due to the fact that V2250 has 1X PCI I/O Backplane (32 bit and 33 Mhz). The V2500 has a 2X PCI I/O Backplane (64 bit, 33Mhz).

Multiple adapter configuration test results

For a Classical IP interface with MTU size of 9180 bytes, the TCP throughput in outbound direction varies almost linearly with the number of adapters, for up to eight 622 adapters and 12 155 Adapters for V2500 and N4000 platforms. See Figures 2 and 3 below.

Figure 2: TCP Outbound throughput Scalability (N4000: 8 CPUs, 622 Adapters, Classical IP)

Figure 3: TCP throughput Scalability (V2500: 8 CPUs, 155 Adapters, Classical IP)

Each netperf process was bound to the same CPU that the adapter uses to send interrupts to. The Classical IP interface on each adapter was configured with a different IP subnet; that is, there were eight IP subnets configured in an eight-adapter test.

On V-series platforms, one 622 adapter was configured per PCI Controller, and a maximum of two 155 adapters were installed in a PCI Controller. On N-series platforms, the 622 Adapters were installed in the Twin Turbo slots.

*8 CPUs, Classical IP

*8 CPUs, Classical IP
² The Inbound and Bidirectional throughput data for the N4000 platform is estimated.

*N4000: 8 CPUs, 4 155 Adapters

*N4000: 8 CPUs, 4 622 Adapters

Tuning guidelines

• System-related Guidelines

• Application-related Guidelines

• Network design-related Guidelines

Appendix A: test methodology and test environment

Test methodology

HP's Netperf version 2.1 was used as the test tool in all the measurements. Netperf measures the throughput at the TCP level. This measurement models many real life applications since these applications use the TCP protocol.

The target systems were configured with a minimum of 2 Gbytes of memory. The ATM software version used was I.11.00.10 with Build Revision 7.39; this corresponds to the DART45 release of ATM software. HP-UX 11.0 was used as the underlying operating system. All the configuration parameters were default with the exception of modifying the LANE MTU size. See Figures 4 and 5 below for the test topology.

Figure 4: single adapter test configuration

The FORE systems ASX-1000 ATM switch ran ForeThought version 4.0 switch software. The Cisco 1010 ATM switch ran version 11.1 of the CISCO IOS software. The N4000 platform had all the four memory carriers installed for the multiple adapter tests. Table 10 below provides the server and workstation platform configurations on which the performance tests were run.

Figure 5: multiple adapter test configuration

³ 8 GB for the Multiple adapter tests.

Protocol overheads and theoretical maximum TCP throughput

The throughput as seen by a TCP application is less than the OC-3 physical layer throughput of 155 Mbps or OC-12 (622 Mbps). This is because of the various protocol headers that are inserted by the protocol stack. For example, the protocol headers for Classical IP interface are as follows:

• TCP/IP Overhead

  This Consists of 20 bytes of TCP header + 20 bytes of IP header + eight bytes of LLC header. This amounts to 48 bytes

• ATM AAL5 Overhead

  The Convergence Sublayer appends an 8-byte trailer. It will also add a Pad. The Pad size varies from 0-47 bytes, depending on Application Message size.

• ATM Layer Overhead

  The ATM Layer will append a five byte header to each Segmentation and Reassembly PDU (48 bytes) to form an ATM cell.

• Physical Layer Overhead

  The Transmission Convergence sublayer creates SONET frames of 27 ATM cells by appending one overhead ATM cell to 26 data ATM cells.

  This leads to the following equation for Theoretical TCP throughput (T)

  T = (D / (D+ 56 + p)) * (48/53) * (26/27) * 155.52 Mbps (or 622 Mbps)

  Where, D = application message size in bytes.

  P = 0-47 bytes padding added at AAL5 layer.

  For Message sizes smaller than 4 Kbytes, The TCP/IP and AAL5 layer protocol overhead forms a major part and theoretical throughput reduces dramatically. For typical message sizes, the theoretical TCP throughput is as follows:

It should be noted that the above calculation assumes Classical IP interface. LAN Emulation adds additional 8 bytes as header. Therefore, the theoretical throughput will decrease accordingly.