From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from smtpng2.m.smailru.net (smtpng2.m.smailru.net [94.100.179.3]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by dev.tarantool.org (Postfix) with ESMTPS id A786D46970F for ; Fri, 29 Nov 2019 20:40:02 +0300 (MSK) Date: Fri, 29 Nov 2019 20:37:54 +0300 From: Igor Munkin Message-ID: <20191129173754.GD1214@tarantool.org> References: <20191126211701.mhavpytwkemux3vm@tkn_work_nb> <20191127083123.GA2752@atlas> <20191128123445.GC29714@atlas> <20191128130005.GA1214@tarantool.org> <20191128131804.GE29714@atlas> <20191128140340.GB1214@tarantool.org> <20191128155801.GB11584@atlas> <20191128183615.GC1214@tarantool.org> <20191129054107.GF15149@atlas> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline In-Reply-To: <20191129054107.GF15149@atlas> Subject: Re: [Tarantool-patches] [PATCH] Move txn from shema to a separate module (use C API instead of FFI) List-Id: Tarantool development patches List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Konstantin Osipov Cc: tarantool-patches@dev.tarantool.org Kostja, On 29.11.19, Konstantin Osipov wrote: > * Igor Munkin [19/11/28 21:39]: > > > Great, we discussed with Leonid and Sasha offline and agreed to make > > several benchmarks to be provided in this thread. With no benchmarks all > > our estimates can be simply wrong. > > For benchmarks, it is always good to refresh them but since you've > been working with LuaJIT for a while you must know some of > the conclusions already: > > * JITed code can be 5-8x faster than Lua + Lua/C > * FFI is 1.5-2x slower than Lua/C in non-JIT mode. So using FFI > is an extra cost if JIT is OFF. > * The way traces are collected, cached and stored has its own > overhead, so JIT ON can be harmful to performance if Lua code > base gets big. Additionally, enabling JIT can nerf performance on a heterogeneous data considering the guarded *LOADs. I can confirm nothing for the numbers you've written above. There are several benchmarks for LuaJIT itself regarding a different workload. My point was about benchmarks for LuaJIT within Tarantool, since we are going to tune it for Tarantool environment primarily. > > In practice, for Tarantool apps, I found all of the above to not > have much impact. The #1 cause of performance slowdown in > Tarantool apps is boxing/unboxing Tarantool tuples and garbage > collection. > > So to measure the real impact of the change, micro-benchmarking > is insufficient. One has to use run e.g. vshard or data > grid or other complex app stress test with wal_mode=off to > see the effect of such change. > > -- > Konstantin Osipov, Moscow, Russia -- Best regards, IM