Forum Discussion
T__Jennings_924
Nimbostratus
NimbostratusPerformance ceiling on LTM 3400?
Have a LTM 3400 on 9.4.4 hosting a handful of virtuals but only one with significant load, which is a FastL4 virtual.
The 3400 is rated on paper for a max of 1gbps of throughput. We're hitting some kind of ceiling:
* the webgui graph shows throughput topping out and flattening at ~725mbps
* the 'b interface' and cacti graphs watching the packet drops increment considerably on both the inside and outside interfaces in the same proportions.
* External monitoring of our application has two well defined cases: fast sub-second response times and slow 3-second response times (consistent with a 3s tcp retry).
* Our provider shows no increment of Qos counters on our link.
The 3400 is rated on paper with a "best case" throughput of 1gbps. I have no idea what's considered "best case".
Is it reasonable to assert that the 3400 is underpowered for this application?
What other metrics could I check?
What else would cause packet drops (reported by 'b interface' but not netstat) on both interfaces in the same proportion? Are interface traffic counters (i.e., read by cacti) pre or post dropped packets?
Does the webgui graph for throughput accurately reflect the capabilities of the device?
Other measurements of interest: the application routinely runs at >200k connections and right around 60k packets/sec in both directions.
A best-case environment has zero packet loss and very low or no network latency. For maximum throughput, a best-case traffic profile would consist of very long-lived connections with very high utilization and virtually no delay in client or server responses. An ideal example might be a corporate, local-access, OS update server only serving large, static files. Obviously not particularly common.
It isn't possible to determine if the device is underpowered without knowing any of the system stats. Here are a couple things I would check.
- What does your CPU utilization look like? Is it peaking, or do you hit your peak throughput before you hit 90% CPU utilization?
- You mentioned 200K connections, how many new connections/sec are you getting?
- Do the /var/log/ltm or /var/log/tmm files have any interesting messages?
--jesse
L4L7_53191Another huge factor in a pure throughput discussion is packet size (usually responses, in particular). Obviously, the larger your payloads, the more throughput you'll get. Have a look at the HTTP profile statistics to get a feel for how large your requests/responses are.
-Matt
