Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Fri, 12 Oct 2007 09:08:58 -0700 Brandeburg, Jesse [EMAIL PROTECTED] wrote: Andi Kleen wrote: When the hw TX queue gains space, the driver self-batches packets from the sw queue to the hw queue. I don't really see the advantage over the qdisc in that scheme. It's certainly not simpler and probably more code and would likely also not require less locks (e.g. a currently lockless driver would need a new lock for its sw queue). Also it is unclear to me it would be really any faster. related to this comment, does Linux have a lockless (using atomics) singly linked list element? That would be very useful in a driver hot path. Use RCU? or write a generic version and get it reviewed. You really want someone with knowledge of all the possible barrier impacts to review it. -- Stephen Hemminger [EMAIL PROTECTED] - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
Use RCU? or write a generic version and get it reviewed. You really want someone with knowledge of all the possible barrier impacts to review it. I guess he was thinking of using cmpxchg; but we don't support this in portable code. RCU is not really suitable for this because it assume writing is relatively rare which is definitely not the case for a qdisc. Also general list management with RCU is quite expensive anyways -- it would require a full copy (that is the 'C' in RCU which Linux generally doesn't use at all) -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
RE: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
Andi Kleen wrote: When the hw TX queue gains space, the driver self-batches packets from the sw queue to the hw queue. I don't really see the advantage over the qdisc in that scheme. It's certainly not simpler and probably more code and would likely also not require less locks (e.g. a currently lockless driver would need a new lock for its sw queue). Also it is unclear to me it would be really any faster. related to this comment, does Linux have a lockless (using atomics) singly linked list element? That would be very useful in a driver hot path. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
related to this comment, does Linux have a lockless (using atomics) singly linked list element? That would be very useful in a driver hot path. No; it doesn't. At least not a portable one. Besides they tend to be not faster anyways because e.g. cmpxchg tends to be as slow as an explicit spinlock. -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
Hi Dave, David Miller wrote on 10/10/2007 02:13:31 AM: Hopefully that new qdisc will just use the TX rings of the hardware directly. They are typically large enough these days. That might avoid some locking in this critical path. Indeed, I also realized last night that for the default qdiscs we do a lot of stupid useless work. If the queue is a FIFO and the device can take packets, we should send it directly and not stick it into the qdisc at all. Since you are talking of how it should be done in the *current* code, I feel LLTX drivers will not work nicely with this. Actually I was trying this change a couple of weeks back, but felt that doin go would result in out of order packets (skbs present in q which were not sent out for LLTX failure will be sent out only at next net_tx_action, while other skbs are sent ahead). One option is to first call qdisc_run() and then process this skb, but that is ugly (requeue handling). However I guess this can be done cleanly once LLTX is removed. Thanks, - KK - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
A 256 entry TX hw queue fills up trivially on 1GB and 10GB, but if you With TSO really? increase the size much more performance starts to go down due to L2 cache thrashing. Another possibility would be to consider using cache avoidance instructions while updating the TX ring (e.g. write combining on x86) -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: Andi Kleen [EMAIL PROTECTED] Date: Wed, 10 Oct 2007 11:16:44 +0200 A 256 entry TX hw queue fills up trivially on 1GB and 10GB, but if you With TSO really? Yes. increase the size much more performance starts to go down due to L2 cache thrashing. Another possibility would be to consider using cache avoidance instructions while updating the TX ring (e.g. write combining on x86) The chip I was working with at the time (UltraSPARC-IIi) compressed all the linear stores into 64-byte full cacheline transactions via the store buffer. It's true that it would allocate in the L2 cache on a miss, which is different from your suggestion. In fact, such a thing might not pan out well, because most of the time you write a single descriptor or two, and that isn't a full cacheline, which means a read/modify/write is the only coherent way to make such a write to RAM. Sure you could batch, but I'd rather give the chip work to do unless I unequivocably knew I'd have enough pending to fill a cacheline's worth of descriptors. And since you suggest we shouldn't queue in software... :-) - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Wed, Oct 10, 2007 at 11:16:44AM +0200, Andi Kleen wrote: A 256 entry TX hw queue fills up trivially on 1GB and 10GB, but if you With TSO really? Hardware queues are generally per-page rather than per-skb so it'd fill up quicker than a software queue even with TSO. Cheers, -- Visit Openswan at http://www.openswan.org/ Email: Herbert Xu ~{PmVHI~} [EMAIL PROTECTED] Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Wed, Oct 10, 2007 at 02:25:50AM -0700, David Miller wrote: The chip I was working with at the time (UltraSPARC-IIi) compressed all the linear stores into 64-byte full cacheline transactions via the store buffer. That's a pretty old CPU. Conclusions on more modern ones might be different. In fact, such a thing might not pan out well, because most of the time you write a single descriptor or two, and that isn't a full cacheline, which means a read/modify/write is the only coherent way to make such a write to RAM. x86 WC does R-M-W and is coherent of course. The main difference is just that the result is not cached. When the hardware accesses the cache line then the cache should be also invalidated. Sure you could batch, but I'd rather give the chip work to do unless I unequivocably knew I'd have enough pending to fill a cacheline's worth of descriptors. And since you suggest we shouldn't queue in software... :-) Hmm, it probably would need to be coupled with batched submission if multiple packets are available you're right. Probably not worth doing explicit queueing though. I suppose it would be an interesting experiment at least. -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: Andi Kleen [EMAIL PROTECTED] Date: Wed, 10 Oct 2007 12:23:31 +0200 On Wed, Oct 10, 2007 at 02:25:50AM -0700, David Miller wrote: The chip I was working with at the time (UltraSPARC-IIi) compressed all the linear stores into 64-byte full cacheline transactions via the store buffer. That's a pretty old CPU. Conclusions on more modern ones might be different. Cache matters, just scale the numbers. I suppose it would be an interesting experiment at least. Absolutely. I've always gotten very poor results when increasing the TX queue a lot, for example with NIU the point of diminishing returns seems to be in the range of 256-512 TX descriptor entries and this was with 1.6Ghz cpus. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
We've done similar testing with ixgbe to push maximum descriptor counts, and we lost performance very quickly in the same range you're quoting on NIU. Did you try it with WC writes to the ring or CLFLUSH? -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Tue, 09 Oct 2007, David Miller wrote: From: jamal [EMAIL PROTECTED] Date: Tue, 09 Oct 2007 17:56:46 -0400 if the h/ware queues are full because of link pressure etc, you drop. We drop today when the s/ware queues are full. The driver txmit lock takes place of the qdisc queue lock etc. I am assuming there is still need for that locking. The filter/classification scheme still works as is and select classes which map to rings. tc still works as is etc. I understand your suggestion. We have to keep in mind, however, that the sw queue right now is 1000 packets. I heavily discourage any driver author to try and use any single TX queue of that size. Which means that just dropping on back pressure might not work so well. Or it might be perfect and signal TCP to backoff, who knows! :-) I can't remember the details anymore, but for 10-GigE, I have encountered cases where I was able to significantly increase TCP performance by increasing the txqueuelen to 1, which is the setting I now use for any 10-GigE testing. -Bill - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
RE: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
-Original Message- From: Andi Kleen [mailto:[EMAIL PROTECTED] Sent: Wednesday, October 10, 2007 9:02 AM To: Waskiewicz Jr, Peter P Cc: David Miller; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; netdev@vger.kernel.org; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED] Subject: Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching We've done similar testing with ixgbe to push maximum descriptor counts, and we lost performance very quickly in the same range you're quoting on NIU. Did you try it with WC writes to the ring or CLFLUSH? -Andi Hmm, I think it might be slightly different, but it still shows queue depth vs. performance. I was actually referring to how many descriptors we can represent a packet with before it becomes a problem wrt performance. This morning I tried to actually push my ixgbe NIC hard enough to come close to filling the ring with packets (384-byte packets), and even on my 8-core Xeon I can't do it. My system can't generate enough I/O to fill the hardware queues before CPUs max out. -PJ Waskiewicz - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: jamal [EMAIL PROTECTED] Date: Wed, 10 Oct 2007 09:08:48 -0400 On Wed, 2007-10-10 at 03:44 -0700, David Miller wrote: I've always gotten very poor results when increasing the TX queue a lot, for example with NIU the point of diminishing returns seems to be in the range of 256-512 TX descriptor entries and this was with 1.6Ghz cpus. Is it interupt per packet? From my experience, you may find interesting results varying tx interupt mitigation parameters in addition to the ring parameters. Unfortunately when you do that, optimal parameters also depends on packet size. so what may work for 64B, wont work well for 1400B. No, it was not interrupt per-packet, I was telling the chip to interrupt me every 1/4 of the ring. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: Bill Fink [EMAIL PROTECTED] Date: Wed, 10 Oct 2007 12:02:15 -0400 On Tue, 09 Oct 2007, David Miller wrote: We have to keep in mind, however, that the sw queue right now is 1000 packets. I heavily discourage any driver author to try and use any single TX queue of that size. Which means that just dropping on back pressure might not work so well. Or it might be perfect and signal TCP to backoff, who knows! :-) I can't remember the details anymore, but for 10-GigE, I have encountered cases where I was able to significantly increase TCP performance by increasing the txqueuelen to 1, which is the setting I now use for any 10-GigE testing. For some reason this does not surprise me. We bumped the ethernet default up to 1000 for gigabit. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On 09 Oct 2007 18:51:51 +0200 Andi Kleen [EMAIL PROTECTED] wrote: David Miller [EMAIL PROTECTED] writes: 2) Switch the default qdisc away from pfifo_fast to a new DRR fifo with load balancing using the code in #1. I think this is kind of in the territory of what Peter said he is working on. Hopefully that new qdisc will just use the TX rings of the hardware directly. They are typically large enough these days. That might avoid some locking in this critical path. I know this is controversial, but realistically I doubt users benefit at all from the prioritization that pfifo provides. I agree. For most interfaces the priority is probably dubious. Even for DSL the prioritization will be likely usually done in a router these days. Also for the fast interfaces where we do TSO priority doesn't work very well anyways -- with large packets there is not too much to prioritize. 3) Work on discovering a way to make the locking on transmit as localized to the current thread of execution as possible. Things like RCU and statistic replication, techniques we use widely elsewhere in the stack, begin to come to mind. If the data is just passed on to the hardware queue, why is any locking needed at all? (except for the driver locking of course) -Andi I wonder about the whole idea of queueing in general at such high speeds. Given the normal bi-modal distribution of packets, and the predominance of 1500 byte MTU; does it make sense to even have any queueing in software at all? -- Stephen Hemminger [EMAIL PROTECTED] - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
I wonder about the whole idea of queueing in general at such high speeds. Given the normal bi-modal distribution of packets, and the predominance of 1500 byte MTU; does it make sense to even have any queueing in software at all? Yes that is my point -- it should just pass it through directly and the driver can then put it into the different per CPU (or per whatever) queues managed by the hardware. The only thing the qdisc needs to do is to set some bit that says it is ok to put this into difference queues; don't need strict ordering Otherwise if the drivers did that unconditionally they might cause problems with other qdiscs. This would also require that the driver exports some hint to the upper layer on how large its internal queues are. A device with a short queue would still require pfifo_fast. Long queue devices could just pass through. That again could be a single flag. -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: Andi Kleen [EMAIL PROTECTED] Date: 09 Oct 2007 18:51:51 +0200 Hopefully that new qdisc will just use the TX rings of the hardware directly. They are typically large enough these days. That might avoid some locking in this critical path. Indeed, I also realized last night that for the default qdiscs we do a lot of stupid useless work. If the queue is a FIFO and the device can take packets, we should send it directly and not stick it into the qdisc at all. If the data is just passed on to the hardware queue, why is any locking needed at all? (except for the driver locking of course) Absolutely. Our packet scheduler subsystem is great, but by default it should just get out of the way. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Tue, 09 Oct 2007 13:43:31 -0700 (PDT) David Miller [EMAIL PROTECTED] wrote: From: Andi Kleen [EMAIL PROTECTED] Date: 09 Oct 2007 18:51:51 +0200 Hopefully that new qdisc will just use the TX rings of the hardware directly. They are typically large enough these days. That might avoid some locking in this critical path. Indeed, I also realized last night that for the default qdiscs we do a lot of stupid useless work. If the queue is a FIFO and the device can take packets, we should send it directly and not stick it into the qdisc at all. If the data is just passed on to the hardware queue, why is any locking needed at all? (except for the driver locking of course) Absolutely. Our packet scheduler subsystem is great, but by default it should just get out of the way. I was thinking why not have a default transmit queue len of 0 like the virtual devices. -- Stephen Hemminger [EMAIL PROTECTED] - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: Stephen Hemminger [EMAIL PROTECTED] Date: Tue, 9 Oct 2007 13:53:40 -0700 I was thinking why not have a default transmit queue len of 0 like the virtual devices. I'm not so sure. Even if the device has huge queues I still think we need a software queue for when the hardware one backs up. It is even beneficial to stick with reasonably sized TX queues because it keeps the total resident state accessed by the CPU within the bounds of the L2 cache. If you go past that it actually hurts to make the TX queue larger instead of helps even if it means you never hit back pressure. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Tue, 2007-09-10 at 14:22 -0700, David Miller wrote: Even if the device has huge queues I still think we need a software queue for when the hardware one backs up. It should be fine to just pretend the qdisc exists despite it sitting in the driver and not have s/ware queues at all to avoid all the challenges that qdiscs bring; if the h/ware queues are full because of link pressure etc, you drop. We drop today when the s/ware queues are full. The driver txmit lock takes place of the qdisc queue lock etc. I am assuming there is still need for that locking. The filter/classification scheme still works as is and select classes which map to rings. tc still works as is etc. cheers, jamal - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: jamal [EMAIL PROTECTED] Date: Tue, 09 Oct 2007 17:56:46 -0400 if the h/ware queues are full because of link pressure etc, you drop. We drop today when the s/ware queues are full. The driver txmit lock takes place of the qdisc queue lock etc. I am assuming there is still need for that locking. The filter/classification scheme still works as is and select classes which map to rings. tc still works as is etc. I understand your suggestion. We have to keep in mind, however, that the sw queue right now is 1000 packets. I heavily discourage any driver author to try and use any single TX queue of that size. Which means that just dropping on back pressure might not work so well. Or it might be perfect and signal TCP to backoff, who knows! :-) While working out this issue in my mind, it occured to me that we can put the sw queue into the driver as well. The idea is that the network stack, as in the pure hw queue scheme, unconditionally always submits new packets to the driver. Therefore even if the hw TX queue is full, the driver can still queue to an internal sw queue with some limit (say 1000 for ethernet, as is used now). When the hw TX queue gains space, the driver self-batches packets from the sw queue to the hw queue. It sort of obviates the need for mid-level queue batching in the generic networking. Compared to letting the driver self-batch, the mid-level batching approach is pure overhead. We seem to be sort of all mentioning similar ideas. For example, you can get the above kind of scheme today by using a mid-level queue length of zero, and I believe this idea was mentioned by Stephen Hemminger earlier. I may experiment with this in the NIU driver. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
On Tue, Oct 09, 2007 at 05:04:35PM -0700, David Miller wrote: We have to keep in mind, however, that the sw queue right now is 1000 packets. I heavily discourage any driver author to try and use any single TX queue of that size. Why would you discourage them? If 1000 is ok for a software queue why would it not be ok for a hardware queue? Which means that just dropping on back pressure might not work so well. Or it might be perfect and signal TCP to backoff, who knows! :-) 1000 packets is a lot. I don't have hard data, but gut feeling is less would also do. And if the hw queues are not enough a better scheme might be to just manage this in the sockets in sendmsg. e.g. provide a wait queue that drivers can wake up and let them block on more queue. The idea is that the network stack, as in the pure hw queue scheme, unconditionally always submits new packets to the driver. Therefore even if the hw TX queue is full, the driver can still queue to an internal sw queue with some limit (say 1000 for ethernet, as is used now). When the hw TX queue gains space, the driver self-batches packets from the sw queue to the hw queue. I don't really see the advantage over the qdisc in that scheme. It's certainly not simpler and probably more code and would likely also not require less locks (e.g. a currently lockless driver would need a new lock for its sw queue). Also it is unclear to me it would be really any faster. -Andi - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [ofa-general] Re: [PATCH 2/3][NET_BATCH] net core use batching
From: Andi Kleen [EMAIL PROTECTED] Date: Wed, 10 Oct 2007 02:37:16 +0200 On Tue, Oct 09, 2007 at 05:04:35PM -0700, David Miller wrote: We have to keep in mind, however, that the sw queue right now is 1000 packets. I heavily discourage any driver author to try and use any single TX queue of that size. Why would you discourage them? If 1000 is ok for a software queue why would it not be ok for a hardware queue? Because with the software queue, you aren't accessing 1000 slots shared with the hardware device which does shared-ownership transactions on those L2 cache lines with the cpu. Long ago I did a test on gigabit on a cpu with only 256K of L2 cache. Using a smaller TX queue make things go faster, and it's exactly because of these L2 cache effects. 1000 packets is a lot. I don't have hard data, but gut feeling is less would also do. I'll try to see how backlogged my 10Gb tests get when a strong sender is sending to a weak receiver. And if the hw queues are not enough a better scheme might be to just manage this in the sockets in sendmsg. e.g. provide a wait queue that drivers can wake up and let them block on more queue. TCP does this already, but it operates in a lossy manner. I don't really see the advantage over the qdisc in that scheme. It's certainly not simpler and probably more code and would likely also not require less locks (e.g. a currently lockless driver would need a new lock for its sw queue). Also it is unclear to me it would be really any faster. You still need a lock to guard hw TX enqueue from hw TX reclaim. A 256 entry TX hw queue fills up trivially on 1GB and 10GB, but if you increase the size much more performance starts to go down due to L2 cache thrashing. - To unsubscribe from this list: send the line unsubscribe netdev in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html