Messages in this thread | | | Subject | Re: [PATCH 1/3] Kernel interfaces for multiqueue aware socket | From | Eric Dumazet <> | Date | Thu, 16 Dec 2010 05:44:20 +0100 |
| |
Le mercredi 15 décembre 2010 à 17:14 -0800, Fenghua Yu a écrit : > On Wed, Dec 15, 2010 at 12:48:38PM -0800, Eric Dumazet wrote: > > Le mercredi 15 décembre 2010 à 12:02 -0800, Fenghua Yu a écrit : > > > From: Fenghua Yu <fenghua.yu@intel.com> > > > > > > Multiqueue and multicore provide packet parallel processing methodology. > > > Current kernel and network drivers place one queue on one core. But the higher > > > level socket doesn't know multiqueue. Current socket only can receive or send > > > packets through one network interfaces. In some cases e.g. multi bpf filter > > > tcpdump and snort, a lot of contentions come from socket operations like ring > > > buffer. Even if the application itself has been fully parallelized and run on > > > multi-core systems and NIC handlex tx/rx in multiqueue in parallel, network layer > > > and NIC device driver assemble packets to a single, serialized queue. Thus the > > > application cannot actually run in parallel in high speed. > > > > > > To break the serialized packets assembling bottleneck in kernel, one way is to > > > allow socket to know multiqueue associated with a NIC interface. So each socket > > > can handle tx/rx in one queue in parallel. > > > > > > Kernel provides several interfaces by which sockets can be bound to rx/tx queues. > > > User applications can configure socket by providing several sockets that each > > > bound to a single queue, applications can get data from kernel in parallel. After > > > that, competitions mentioned above can be removed. > > > > > > With this patch, the user-space receiving speed on a Intel SR1690 server with > > > a single L5640 6-core processor and a single ixgbe-based NIC goes from 0.73Mpps > > > to 4.20Mpps, nearly a linear speedup. A Intel SR1625 server two E5530 4-core > > > processors and a single ixgbe-based NIC goes from 0.80Mpps to 4.6Mpps. We noticed > > > the performance penalty comes from NUMA memory allocation. > > > > > > > ??? please elaborate on these NUMA memory allocations. This should be OK > > after commit 564824b0c52c34692d (net: allocate skbs on local node) > >
No data for this NUMA problem ? We had to convince Andrew Morton for this patch to get in.
> > > This patch set provides kernel ioctl interfaces for user space. User space can > > > either directly call the interfaces or libpcap interfaces can be further provided > > > on the top of the kernel ioctl interfaces. > > > > So, say we have 8 queues, you want libpcap opens 8 sockets, and bind > > them to each queue. Add a bpf filter to each one of them. This seems not > > generic way, because it wont work for an UDP socket for example. > > This only works for AF_PACKET like this patch set shows. >
Yes, we also should address other sockets, with generic mechanisms.
> > And you already can do this using SKF_AD_QUEUE (added in commit > > d19742fb) > > SKF_AD_QUEUE doesn't know number of rx queues. Thus user application can't > specify right SKF_AD_QUEUE. > > SKF_AD_QUEUE only works for rx. There is no queue bound interfaces for tx. > > I can change the patch set to use SKF_AD_QUEUE by removing the set rx queue > interface and still keep interfaces of > #define SIOGNUMRXQUEUE 0x8939 /* Get number of rx queues. */ > #define SIOGNUMTXQUEUE 0x893A /* Get number of tx queues. */ > #define SIOSTXQUEUEMAPPING 0x893C /* Set tx queue mapping. */ > #define SIOGRXQUEUEMAPPING 0x893D /* Get rx queue mapping. */ > #define SIOGTXQUEUEMAPPING 0x893E /* Get tx queue mapping. */ > > > > > Also your AF_PACKET patch only address mmaped sockets. > > > The new patch set will use SKF_AD_QUEUE for rx. So it won't be limited to mmaped > sockets. >
We really need to be smarter than that, not adding raw API.
Tom Herbert added RPS, RFS, XPS, in a way applications dont have to use special API, just run normal code.
Please understand that using 8 AF_PACKET sockets bound to a given device is a total waste, because the way we loop on ptype_all before entering AF_PACKET code, and in 12% of the cases deliver the packet into a queue, and 77.5% of the case reject the packet.
This is absolutely not scalable to say... 64 queues.
I do believe we can handle that using one AF_PACKET socket for the RX side, in order to not slow down the loop we have in __netif_receive_skb()
list_for_each_entry_rcu(ptype, &ptype_all, list) { ... deliver_skb(skb, pt_prev, orig_dev); }
(Same problem with dev_queue_xmit_nit() by the way, even worse since we skb_clone() packet _before_ entering af_packet code)
And we can change af_packet to split the load to N skb queues or N ring buffers, N not being necessarly number of NIC queues, but the number needed to handle the expected load.
There is nothing preventing us changing af_packet/udp/tcp_listener to something more scalable in itself, using a set of receive queues, and NUMA friendly data set. We did multiqueue for a net_device like this, not adding N pseudo devices as we could have done.
-- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |