Re: [nsp] 7505 / VIP 4-80 memory query

From: Hank Nussbacher (hank@att.net.il)
Date: Wed Feb 27 2002 - 01:25:55 EST


At 02:37 PM 26-02-02 +0000, Thomas, Iwan wrote:

We did some benchmarks on this as well.

Packet size pps Theoretical incoming bandwidth Incoming
Bandwidth Outgoing Bandwidth PBR CPU/VIP load
64 150000 77Mb/sec 65Mb/sec
     65Mb/sec OFF 0%/82%
64 200000 102Mb/sec 90Mb/sec
     90Mb/sec OFF 0%/92%
64 240000 123Mb/sec 107Mb/sec
                             OFF 0%/99%
256 200000 408Mb/sec 350Mb/sec
     350Mb/sec OFF 0%/79%
64 200000 102Mb/sec 70Mb/sec
     60Mb/sec ON 0%/99%/97%
256 200000 408Mb/sec 350Mb/sec
     326Mb/sec ON 0%/99%/97%

Notes:
"PC Setup: Incoming packets were generated by a Pentium IV 1.4Ghz,
with 128MB and an Intel Pro-1000F NIC card, sending SYN
packets"

"Cisco setup: Cisco 7513, IOS 12.0(18)S, RSP4, 256MB, GEIP+ (VIP4-80) cards,
running with "ip cef", "ip cef distributed", "ip route-cache
flow"
PBR setup: 50% of incoming data was intended to be
PBRed
In addition, CSCdp78100 states "Policy routing does not work in 12.0(8)S
if DCEF is enabled. Workaround is to disable Distributed CEF. Policy
routing does work with CEF." The bugid does not list a fixed version.

Bottom line: the VIP CPU burns out and can't handle the load. You need to
check the VIP CPU which is the bottleneck.

-Hank Nussbacher
Consultant
Wanwall Ltd.

>Hi folks,
>
>I'm stress-testing a 7505's throughput with a QoS service-policy applied
>outbound on a POS STM-1 interface. I am seeing severe degradation of
>throughput when compared to FIFO w/ no policy. Distributed CEF is on the
>box, route-cache-distributed on the relevant interfaces.
>
>With no policy applied, I can generate >100Mbps full-duplex, using 64-byte
>packets (giving me >200,000 pps in each direction).
>With the policy applied, I drop to 33Mbps, <70,000 pps. Note that no packets
>are being queued or dropped according to my CBWFQ stats.
>
>I'm trying to pin down where the bottleneck is most likely caused, without
>going out and sourcing extra memory ;-) The VIP CPU memory is responsible
>for implementing policing and assigning packets the correct DSCP marker (I
>think). So am I right in guessing that an upgrade to 128MB would improve the
>siuation, or am I missing something? Either way, a little advice/data on
>the responsibilities of the separate memory portions of the VIPs and the RSP
>would be appreciated.
>
>Specs:
>
>RSP4 (R5000 CPU at 200MHz) with 128MB DRAM and 2MB SDRAM.
>2 x VIP 4-80 RM7000 with 64MB CPU SDRAM and 64MB Packet SDRAM.
>2 x POS PA-POSSW-SM port adapters.
>The POS links are on separate VIPs, running STM-1 direct to a traffic
>generator.
>
>Cheers,
>
>Iwan
>
>
>**********************************************************************
>This message may contain information which is confidential or privileged.
>If you are not the intended recipient, please advise the sender immediately
>by reply e-mail and delete this message and any attachments
>without retaining a copy.
>
>**********************************************************************



This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:13:34 EDT