From mboxrd@z Thu Jan 1 00:00:00 1970 From: Evgeniy Polyakov Subject: Re: [take13 1/3] kevent: Core files. Date: Fri, 25 Aug 2006 11:20:08 +0400 Message-ID: <20060825072007.GE1816@2ka.mipt.ru> References: <11563322941645@2ka.mipt.ru> <11563322971212@2ka.mipt.ru> <20060824200322.GA19533@infradead.org> <20060825054815.GC16504@2ka.mipt.ru> <20060824232024.0d230823.akpm@osdl.org> <20060825063238.GD16504@2ka.mipt.ru> <20060824235859.f8840fb2.akpm@osdl.org> Mime-Version: 1.0 Content-Type: text/plain; charset=koi8-r Cc: Christoph Hellwig , lkml , David Miller , Ulrich Drepper , netdev , Zach Brown Return-path: To: Andrew Morton Content-Disposition: inline In-Reply-To: <20060824235859.f8840fb2.akpm@osdl.org> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Thu, Aug 24, 2006 at 11:58:59PM -0700, Andrew Morton (akpm@osdl.org) wrote: > > > > kmalloc is really slow actually - it always shows somewhere on top > > > > in profiles and brings noticeble overhead > > > > > > It shouldn't. Please describe the workload and send the profiles. > > > > epoll based trivial server (accept + sendfile for the same file, about > > 4k), httperf with big amount of simulateneous connections. 3c59x NIC > > (with e1000 there were no ioreads and netif_rx). > > __alloc_skb calls kmem_cache_alloc() and ___kmalloc(). > > > > 16158 1.3681 ioread16 > > 8073 0.6835 ioread32 > > 3485 0.2951 irq_entries_start > > 3018 0.2555 _spin_lock > > 2103 0.1781 tcp_v4_rcv > > 1503 0.1273 sysenter_past_esp > > 1492 0.1263 netif_rx > > 1459 0.1235 skb_copy_bits > > 1422 0.1204 _spin_lock_irqsave > > 1145 0.0969 ip_route_input > > 983 0.0832 kmem_cache_free > > 964 0.0816 __alloc_skb > > 926 0.0784 common_interrupt > > 891 0.0754 __do_IRQ > > 846 0.0716 _read_lock > > 826 0.0699 __netif_rx_schedule > > 806 0.0682 __kmalloc > > 767 0.0649 do_tcp_sendpages > > 747 0.0632 __copy_to_user_ll > > 744 0.0630 pskb_expand_head > > > > That doesn't look too bad. > > What's that as a percentage of total user+system time? With e1000 allocations take more time than actual TCP processing, so it rised some suspicious for me (especially in bulk transfer). Total time is about 7 times more than system one, user time is much less than system one (about 20 times less, but test duration was not too long, so it can vary). I do not say it is bad, but it is noticeble and should be eliminated if there are no requirements to have it. -- Evgeniy Polyakov