Dl380 G5 NC510C - speed issue

[KarloChacon]

hi guys

I got a Dl380 g5 - W2003 x64bits - NC510C
When transmitting files like 10 - 15 GB size it takes long time to transmit the files almost 3 minutes

latest driver - firmware same issue
I've disabled RSS - Large Send Offload
but still not the speed we need

these are the NIC values

802.1p QoS Packet Tagging = 0
Maximum Frame Size = 8014 Bytes
Flow Control = 1
Tx IP Checksum Offload = 1
Rx IP Checksum Offload = 1
Tx TCP Checksum Offload = 1
Rx TCP Checksum Offload = 1
Large Send Offload = 0
Number of Tx Descriptors = 1024
Number of Rx Descriptors = 8192
Number of Rx Descriptors (Jumbo) = 4096
TaskOffloadLSOMaxOffloadSize = 64000
TaskOffloadLSOMinSegmentCount = 2
TaskOffloadLSOIPOptions = 0
TaskOffloadLSOTCPOptions = 1
TaskOffloadChecksumIPv4TxIPOptions = 0
TaskOffloadChecksumIPv4RxIPOptions = 0
TaskOffloadChecksumIPv4TxTCPOptions = 0
TaskOffloadChecksumIPv4RxTCPOptions = 0
Receive-Side Scaling = 0


any idea what could it be?
is there any table performance comparison to check if my NIC is OK? I mean maybe you guys are getting better throughput that my server

thanks a lot

[KarloChacon]

anyone?

[rick jones · Accepted solution]

Check the loading of individual CPUs on either end of the transfer. A single TCP connection cannot make use of the services of more than one core (essentially).

I see you already have the MTU maxed. Is the MTU at the _other_ end also up at 8000 bytes? If not, all TCP connections to/from that remote system will still use the smaller MTU.

Make sure there are no TCP retransmissions during the run.

Are you certain the NC510 is in an x8 slot?

Attached is a Linux-centric (my work is more Linux and Unix) bit of boiler-plate I'm working-on - some of the specifics may not apply, but the concepts should.

[KarloChacon]

thanks a lot

I will work with my EU what you recommended

[KarloChacon]

â Check the loading of individual CPUs on either end of the transfer. A single TCP connection cannot make use of the services of more than one core (essentially).

When coping CPU util rises from 0% to 10-11% on sender side and to 6-7% on receiving side.


I see you already have the MTU maxed. Is the MTU at the _other_ end also up at 8000 bytes? If not, all TCP connections to/from that remote system will still use the smaller MTU.
Yes both sides are maxed out on 8000 bytes.


Make sure there are no TCP retransmissions during the run.
I need to collect some TCP packet to tell you that â but I would say that probably there are some retransmission happeningâ ¦



Are you certain the NC510 is in an x8 slot?â â
It is running on x8 slot.

[rick jones]

Just to be certain, only 10% on any _one_ core on either side, or 10% overall? Frankly, 10% on only one core even with JumboFrames going feels low for the receiving side - although I'm not terribly versed in using the stateful TCP offloads if those are being used.