On Sat, Oct 15, 2022 at 12:49:59PM -0700, Cong Wang wrote: > On Mon, Oct 03, 2022 at 12:11:39AM +0000, Bobby Eshleman wrote: > > On Thu, Oct 06, 2022 at 09:34:10AM +0200, Stefano Garzarella wrote: > > > On Thu, Oct 06, 2022 at 03:08:12AM -0400, Michael S. Tsirkin wrote: > > > > On Wed, Oct 05, 2022 at 06:19:44PM -0700, Bobby Eshleman wrote: > > > > > This patch replaces the struct virtio_vsock_pkt with struct sk_buff. > > > > > > > > > > Using sk_buff in vsock benefits it by a) allowing vsock to be extended > > > > > for socket-related features like sockmap, b) vsock may in the future > > > > > use other sk_buff-dependent kernel capabilities, and c) vsock shares > > > > > commonality with other socket types. > > > > > > > > > > This patch is taken from the original series found here: > > > > > https://lore.kernel.org/all/cover.1660362668.git.bobby.eshleman@xxxxxxxxxxxxx/ > > > > > > > > > > Small-sized packet throughput improved by ~5% (from 18.53 Mb/s to 19.51 > > > > > Mb/s). Tested using uperf, 16B payloads, 64 threads, 100s, averaged from > > > > > 10 test runs (n=10). This improvement is likely due to packet merging. > > > > > > > > > > Large-sized packet throughput decreases ~9% (from 27.25 Gb/s to 25.04 > > > > > Gb/s). Tested using uperf, 64KB payloads, 64 threads, 100s, averaged > > > > > from 10 test runs (n=10). > > > > > > > > > > Medium-sized packet throughput decreases ~5% (from 4.0 Gb/s to 3.81 > > > > > Gb/s). Tested using uperf, 4k to 8k payload sizes picked randomly > > > > > according to normal distribution, 64 threads, 100s, averaged from 10 > > > > > test runs (n=10). > > > > > > > > It is surprizing to me that the original vsock code managed to outperform > > > > the new one, given that to my knowledge we did not focus on optimizing it. > > > > > > Yeah mee to. > > > > > > > Indeed. > > > > > From this numbers maybe the allocation cost has been reduced as it performs > > > better with small packets. But with medium to large packets we perform > > > worse, perhaps because previously we were allocating a contiguous buffer up > > > to 64k? > > > Instead alloc_skb() could allocate non-contiguous pages ? (which would solve > > > the problems we saw a few days ago) > > > > > > > I think this would be the case with alloc_skb_with_frags(), but > > internally alloc_skb() uses kmalloc() for the payload and sk_buff_head > > slab allocations for the sk_buff itself (all the more confusing to me, > > as the prior allocator also uses two separate allocations per packet). > > I think it is related to your implementation of > virtio_transport_add_to_queue(), where you introduced much more > complicated logic than before: > > - spin_lock_bh(&vsock->send_pkt_list_lock); > - list_add_tail(&pkt->list, &vsock->send_pkt_list); > - spin_unlock_bh(&vsock->send_pkt_list_lock); > - > + virtio_transport_add_to_queue(&vsock->send_pkt_queue, skb); > I wish it were that easy, but I included this change because it actually boosts performance. For 16B payloads, this change improves throughput from 16 Mb/s to 20Mb/s in my test harness, and reduces the memory usage of the kmalloc-512 and skbuff_head_cache slab caches by ~50MB at cache size peak (total slab cache size from ~540MB to ~390MB), but typically (not at peak) the slab cache size when this merging is used keeps the memory slab caches closer to ~150MB smaller. Tests done using uperf. For payloads greater than GOOD_COPY_LEN I don't see any any notable difference between the skb code with merging and the skb code without merging in terms of throughput. I assume this is because the skb->len comparison with GOOD_COPY_LEN should short circuit the expression and the other memory operations should not occur. > A simple list_add_tail() is definitely faster than your > virtio_transport_skbs_can_merge() check. So, why do you have to merge > skb while we don't merge virtio_vsock_pkt? > sk_buff is over twice the size of virtio_vsock_pkt (96B vs 232B). It seems wise to reduce the footprint in other ways to try and keep it comparable. > _If_ you are trying to mimic TCP, I think you are doing it wrong, it can > be much more efficient if you could do the merge in sendmsg() before skb > is even allocated, see tcp_sendmsg_locked(). I'll definitely give it a read, merging before allocating an skb sounds better. Best, Bobby
