Re: 100Gb performance

From: Daniel Braniss <danny_at_cs.huji.ac.il>
Date: Thu, 19 Jun 2025 13:12:51 UTC
Hi Olivier
The problem I have it’s that one end is an iscilon storage, not much I can do there.
We are running several rsyncs in parallel, getting some better throughput, but
nowhere near the 10GB/, actually just barely touching 1GB/s

Thanks,
	Danny

> On 19 Jun 2025, at 14:55, Olivier Cochard-Labbé <olivier@freebsd.org> wrote:
> 
> 
> On Thu, Jun 19, 2025 at 8:20 AM Daniel Braniss <danny@cs.huji.ac.il <mailto:danny@cs.huji.ac.il>> wrote:
>> hi,
>> 
>> i am running 14.2 on a DELL PowerEdge R750 with a mellanox/nvidia 100Gb nic  mlx5en: 
>> 
>> mce0: flags=1008843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST,LOWER_UP> metric 0 mtu 1500
>>  
>> 	options=66ef07bb<RXCSUM,TXCSUM,VLAN_MTU,VLAN_HWTAGGING,JUMBO_MTU,VLAN_HWCSUM,TSO4,TSO6,LRO,VLAN_HWFILTER,NV,VLAN_HWTSO,LINKSTATE,RXCSUM_IPV6,TXCSUM_IPV6,HWSTATS,HWRXTSTMP,MEXTPG,VXLAN_HWCSUM,VXLAN_HWTSO>
>>        ether ...
>>        inet ... netmask 0xfffffc00 broadcast ….
>>        media: Ethernet 100GBase-KR4 <full-duplex,rxpause,txpause>
>>        status: active
>>        nd6 options=29<PERFORMNUD,IFDISABLED,AUTO_LINKLOCAL>
>> 
>> I’m doing a rsync from an iscsilon mounted via this mce0, the best max throughput is about 1GBs which is a bit depressing
>> 
> 
> Regarding this 1GB/s (8Gb/s) it is how I get on my side with a very simple netcat transfert.
> By simple transfert, I mean using one TCP flow with a single process netcat:
> 
> On the receiver host:
> nc -l 12345 > /dev/null
> On the sender host:
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12345
> 107374182400 bytes transferred in 77.772515 secs (1380618626 bytes/sec)
> 
> Which is about 1.3GB/s, so close to your 1GB/s.
> 
> Let’s dig a little more and one the sender, displaying the stats for each NIC queues.
> How many queues the drivers configured on my sender system ?
> # sysctl dev.mce.0.conf.channels
> dev.mce.0.conf.channels: 40
> So 40 queues (match my output of nproc), great.
> But how many were used during this test:
> # sysctl dev.mce.0 | awk '/txstat.*\.bytes/ && $NF != 0'
> dev.mce.0.txstat26tc0.bytes: 74
> dev.mce.0.txstat20tc0.bytes: 120
> dev.mce.0.txstat4tc0.bytes: 112564632016
> dev.mce.0.txstat0tc0.bytes: 60
> 
> => Only one queue (the number 4 in my example) is used.
> And it is the same problem on the receiver: One queue/one core
> 
> Let’s improving this by running 8 parallels nc at the same time, but we need to use 8 different TCP sessions to let RSS selecting 8 differents queues:
> On the receiver host:
> nc -l 12341 > /dev/null &
> nc -l 12342 > /dev/null &
> nc -l 12343 > /dev/null &
> nc -l 12344 > /dev/null &
> nc -l 12345 > /dev/null &
> nc -l 12346 > /dev/null &
> nc -l 12347 > /dev/null &
> nc -l 12348 > /dev/null
> 
> On the sender host:
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12341 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12342 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12343 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12344 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12345 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12346 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12347 &
> dd if=/dev/zero bs=1G count=100 | nc 1.1.1.30 12348
> 
> Then we need to add the output from all those dd:
> 107374182400 bytes transferred in 103.937552 secs (1033064374 bytes/sec)
> 107374182400 bytes transferred in 104.474689 secs (1027753071 bytes/sec)
> 107374182400 bytes transferred in 104.939627 secs (1023199578 bytes/sec)
> 107374182400 bytes transferred in 105.002306 secs (1022588806 bytes/sec)
> 107374182400 bytes transferred in 105.674894 secs (1016080345 bytes/sec)
> 107374182400 bytes transferred in 105.687319 secs (1015960885 bytes/sec)
> 107374182400 bytes transferred in 106.480994 secs (1008388239 bytes/sec)
> 107374182400 bytes transferred in 106.837954 secs (1005019084 bytes/sec)
> 
> To have a total of 8152054382 bytes/sec (8.15 GBytes/s or 65Gb/s).
> You can check the stats per queue, and you will notice that 8 of them should have been used.
> So you need to use a multi-threaded/parallel rsync equivalent (on both sides) to fill your link.
> 
> 
>> but tcpdump -i mce0 says:
>> store-09# tcpdump -i mce0 host <same net as mce0>
>> tcpdump: verbose output suppressed, use -v[v]... for full protocol decode
>> listening on mce0, link-type EN10MB (Ethernet), snapshot length 262144 bytes
>>                                             **********
>> 
> 
> Don’t worry about the libpcap definition, from contrib/libpcap/pcap/dlt.h :
> #define DLT_NULL        0       /* BSD loopback encapsulation */
> #define DLT_EN10MB      1       /* Ethernet (10Mb) */
> #define DLT_EN3MB       2       /* Experimental Ethernet (3Mb) */
> #define DLT_AX25        3       /* Amateur Radio AX.25 */
> #define DLT_PRONET      4       /* Proteon ProNET Token Ring */
> #define DLT_CHAOS       5       /* Chaos */
> #define DLT_IEEE802     6       /* 802.5 Token Ring */
> #define DLT_ARCNET      7       /* ARCNET, with BSD-style header */
> #define DLT_SLIP        8       /* Serial Line IP */
> #define DLT_PPP         9       /* Point-to-point Protocol */
> #define DLT_FDDI        10      /* FDDI */
> 
> So the EN10MB is simply the term used for "Ethernet".
> 
> Regards,
> Olivier