From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from kanga.kvack.org (kanga.kvack.org [205.233.56.17]) by smtp.lore.kernel.org (Postfix) with ESMTP id 176ADC3DA63 for ; Fri, 26 Jul 2024 07:32:53 +0000 (UTC) Received: by kanga.kvack.org (Postfix) id 8C6156B008C; Fri, 26 Jul 2024 03:32:53 -0400 (EDT) Received: by kanga.kvack.org (Postfix, from userid 40) id 875916B0092; Fri, 26 Jul 2024 03:32:53 -0400 (EDT) X-Delivered-To: int-list-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix, from userid 63042) id 73DD36B0096; Fri, 26 Jul 2024 03:32:53 -0400 (EDT) X-Delivered-To: linux-mm@kvack.org Received: from relay.hostedemail.com (smtprelay0011.hostedemail.com [216.40.44.11]) by kanga.kvack.org (Postfix) with ESMTP id 518B96B008C for ; Fri, 26 Jul 2024 03:32:53 -0400 (EDT) Received: from smtpin27.hostedemail.com (a10.router.float.18 [10.200.18.1]) by unirelay03.hostedemail.com (Postfix) with ESMTP id 055CAA1656 for ; Fri, 26 Jul 2024 07:32:53 +0000 (UTC) X-FDA: 82381087026.27.70D1F2B Received: from mgamail.intel.com (mgamail.intel.com [198.175.65.20]) by imf16.hostedemail.com (Postfix) with ESMTP id A79CC18000B for ; Fri, 26 Jul 2024 07:32:50 +0000 (UTC) Authentication-Results: imf16.hostedemail.com; dkim=pass header.d=intel.com header.s=Intel header.b=B+7rzwkX; spf=pass (imf16.hostedemail.com: domain of ying.huang@intel.com designates 198.175.65.20 as permitted sender) smtp.mailfrom=ying.huang@intel.com; dmarc=pass (policy=none) header.from=intel.com ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=hostedemail.com; s=arc-20220608; t=1721979104; h=from:from:sender:reply-to:subject:subject:date:date: message-id:message-id:to:to:cc:cc:mime-version:mime-version: content-type:content-type: content-transfer-encoding:content-transfer-encoding: in-reply-to:in-reply-to:references:references:dkim-signature; bh=X3yKWNp1Cdi0iLh0gPsETVqTCCxLhM7h0W4bIEFRSUs=; b=V1OlWTMWQ7c2KT/SrINOqo65eOG0eRXD/fUQYAgHrRhiLmywmBi2suQ3AuquPfWzIdAE5h zhabnPwGfGyEukAVs371mshfhuJoMXM4wzrAzrTrolgA3mdJssElAuXDaO4SAWJ0r2OXoW 7a5O5lFGxvNXIyBQky8VevBqlAjyYr8= ARC-Authentication-Results: i=1; imf16.hostedemail.com; dkim=pass header.d=intel.com header.s=Intel header.b=B+7rzwkX; spf=pass (imf16.hostedemail.com: domain of ying.huang@intel.com designates 198.175.65.20 as permitted sender) smtp.mailfrom=ying.huang@intel.com; dmarc=pass (policy=none) header.from=intel.com ARC-Seal: i=1; s=arc-20220608; d=hostedemail.com; t=1721979104; a=rsa-sha256; cv=none; b=2dkATRnGYqFM+P50SSZWkYCUOnxH3UrPTrF+1oG4hQ4Gbz6NaSnvGpAszDUlk+V5bvwVIY mXhF0zj9oH8wM4p7Etzc/FAG+BAnPY4wT3wqGvrdBlOctmSrHb5hU5JP7oFNWBiFr7lJQv WgYI9t45cVFmQrW+M8LO0GzeuJFw5eY= DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1721979171; x=1753515171; h=from:to:cc:subject:in-reply-to:references:date: message-id:mime-version:content-transfer-encoding; bh=xRmILblJ54qqwxCEnXupQO3wlslsHCFAgUEkXfKxiEw=; b=B+7rzwkXcA4MFd+8SmNo5raSV6wW6eqD7jQJ/vzGfSx0M1rI5m8iAdYJ 1O/OtbSq6ulRoH0farM7GHINZ0qLGYO44+Ow2uiD0KQFHxuCEVBZmhrhH efShtTZ+cNTdwq17Y+r4xI2Em/ZZkpHTy/ZLZYgk3B+NWcqp5CUqW94OL 8s0WPxH/K0xSpSvFKelh7BBlP8MCY3hcCL0YRCVwsJObgK3kJ97L7S3mx MR6BlLPUY8usjcUt4gpNktdyiiVXH3hQ1efM1dBeLg0GtYHzVHbHuxeBI OKEA3WdIDIgj5yzxh6OdqEQfhsICCaCCwkZTro5E/kd+UHpBb/rIHmzzX Q==; X-CSE-ConnectionGUID: s6tT+WC9RE+48cX7ZlV4gA== X-CSE-MsgGUID: KF7rJ/P2TNqP0S7VHj//aQ== X-IronPort-AV: E=McAfee;i="6700,10204,11144"; a="19558771" X-IronPort-AV: E=Sophos;i="6.09,238,1716274800"; d="scan'208";a="19558771" Received: from fmviesa006.fm.intel.com ([10.60.135.146]) by orvoesa112.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 26 Jul 2024 00:32:49 -0700 X-CSE-ConnectionGUID: Tn/TOyOzTNqdsijLqqX4vw== X-CSE-MsgGUID: XPWg7DiASZCecmFSS3kVqg== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="6.09,238,1716274800"; d="scan'208";a="52882301" Received: from yhuang6-desk2.sh.intel.com (HELO yhuang6-desk2.ccr.corp.intel.com) ([10.238.208.55]) by fmviesa006-auth.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 26 Jul 2024 00:32:46 -0700 From: "Huang, Ying" To: Chris Li Cc: Andrew Morton , Kairui Song , Ryan Roberts , Kalesh Singh , linux-kernel@vger.kernel.org, linux-mm@kvack.org, Barry Song Subject: Re: [PATCH v3 0/2] mm: swap: mTHP swap allocator base on swap cluster order In-Reply-To: (Chris Li's message of "Fri, 26 Jul 2024 00:22:39 -0700") References: <20240619-swap-allocator-v3-0-e973a3102444@kernel.org> <87v8242vng.fsf@yhuang6-desk2.ccr.corp.intel.com> <87bk3pzr5p.fsf@yhuang6-desk2.ccr.corp.intel.com> <87frrw3aju.fsf@yhuang6-desk2.ccr.corp.intel.com> Date: Fri, 26 Jul 2024 15:29:13 +0800 Message-ID: <877cd8392u.fsf@yhuang6-desk2.ccr.corp.intel.com> User-Agent: Gnus/5.13 (Gnus v5.13) MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable X-Rspamd-Server: rspam06 X-Rspamd-Queue-Id: A79CC18000B X-Stat-Signature: tnfcq8yb7eoia6t3ktaakzot1tcsq6c4 X-Rspam-User: X-HE-Tag: 1721979170-943975 X-HE-Meta: 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 reOk/Ep7 ER0WveNyTr1Y0BTgEpcsK8xGCbzahcQbEY8vM9g04IOnX8xFYKP/OsREYSFWN6TIY5T3lNWDfalSsM0hmcORqD1grcHyNvbdyDHNaOxHQNHXiofgVXL+ILIz+vtvgkReczsu+4QUHb6xLheVPqW+dTaCj8AnKSCKux1G9ZF+SXK+EaSpg55zb5CXrpzjoEi80oMd+3cY3+tkyW/O9SIcB0vmuYi5A97DZIji69JSmvXOovXKxJzfRSj5wGpxm279zi7+7+u2m/bsdMRq4ZnO5BPoMbg== X-Bogosity: Ham, tests=bogofilter, spamicity=0.000000, version=1.2.4 Sender: owner-linux-mm@kvack.org Precedence: bulk X-Loop: owner-majordomo@kvack.org List-ID: List-Subscribe: List-Unsubscribe: Chris Li writes: > On Fri, Jul 26, 2024 at 12:01=E2=80=AFAM Huang, Ying wrote: >> >> Chris Li writes: >> >> > On Mon, Jun 24, 2024 at 7:36=E2=80=AFPM Huang, Ying wrote: >> >> >> >> Chris Li writes: >> >> >> >> > On Wed, Jun 19, 2024 at 7:32=E2=80=AFPM Huang, Ying wrote: >> >> >> >> >> >> Chris Li writes: >> >> >> >> >> >> > This is the short term solutiolns "swap cluster order" listed >> >> >> > in my "Swap Abstraction" discussion slice 8 in the recent >> >> >> > LSF/MM conference. >> >> >> > >> >> >> > When commit 845982eb264bc "mm: swap: allow storage of all mTHP >> >> >> > orders" is introduced, it only allocates the mTHP swap entries >> >> >> > from new empty cluster list. It has a fragmentation issue >> >> >> > reported by Barry. >> >> >> > >> >> >> > https://lore.kernel.org/all/CAGsJ_4zAcJkuW016Cfi6wicRr8N9X+GJJhg= MQdSMp+Ah+NSgNQ@mail.gmail.com/ >> >> >> > >> >> >> > The reason is that all the empty cluster has been exhausted while >> >> >> > there are planty of free swap entries to in the cluster that is >> >> >> > not 100% free. >> >> >> > >> >> >> > Remember the swap allocation order in the cluster. >> >> >> > Keep track of the per order non full cluster list for later allo= cation. >> >> >> > >> >> >> > User impact: For users that allocate and free mix order mTHP swa= pping, >> >> >> > It greatly improves the success rate of the mTHP swap allocation= after the >> >> >> > initial phase. >> >> >> > >> >> >> > Barry provides a test program to show the effect: >> >> >> > https://lore.kernel.org/linux-mm/20240615084714.37499-1-21cnbao@= gmail.com/ >> >> >> > >> >> >> > Without: >> >> >> > $ mthp-swapout >> >> >> > Iteration 1: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 2: swpout inc: 219, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 3: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 4: swpout inc: 219, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 5: swpout inc: 110, swpout fallback inc: 117, Fallback= percentage: 51.54% >> >> >> > Iteration 6: swpout inc: 0, swpout fallback inc: 230, Fallback p= ercentage: 100.00% >> >> >> > Iteration 7: swpout inc: 0, swpout fallback inc: 229, Fallback p= ercentage: 100.00% >> >> >> > Iteration 8: swpout inc: 0, swpout fallback inc: 223, Fallback p= ercentage: 100.00% >> >> >> > Iteration 9: swpout inc: 0, swpout fallback inc: 224, Fallback p= ercentage: 100.00% >> >> >> > Iteration 10: swpout inc: 0, swpout fallback inc: 216, Fallback = percentage: 100.00% >> >> >> > Iteration 11: swpout inc: 0, swpout fallback inc: 212, Fallback = percentage: 100.00% >> >> >> > Iteration 12: swpout inc: 0, swpout fallback inc: 224, Fallback = percentage: 100.00% >> >> >> > Iteration 13: swpout inc: 0, swpout fallback inc: 214, Fallback = percentage: 100.00% >> >> >> > >> >> >> > $ mthp-swapout -s >> >> >> > Iteration 1: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 2: swpout inc: 227, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 3: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 4: swpout inc: 224, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 5: swpout inc: 33, swpout fallback inc: 197, Fallback = percentage: 85.65% >> >> >> > Iteration 6: swpout inc: 0, swpout fallback inc: 229, Fallback p= ercentage: 100.00% >> >> >> > Iteration 7: swpout inc: 0, swpout fallback inc: 223, Fallback p= ercentage: 100.00% >> >> >> > Iteration 8: swpout inc: 0, swpout fallback inc: 219, Fallback p= ercentage: 100.00% >> >> >> > Iteration 9: swpout inc: 0, swpout fallback inc: 212, Fallback p= ercentage: 100.00% >> >> >> > >> >> >> > With: >> >> >> > $ mthp-swapout >> >> >> > Iteration 1: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 2: swpout inc: 219, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 3: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 4: swpout inc: 219, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 5: swpout inc: 227, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 6: swpout inc: 230, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > ... >> >> >> > Iteration 94: swpout inc: 224, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 95: swpout inc: 221, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 96: swpout inc: 229, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 97: swpout inc: 219, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 98: swpout inc: 222, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 99: swpout inc: 223, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 100: swpout inc: 224, swpout fallback inc: 0, Fallback= percentage: 0.00% >> >> >> > >> >> >> > $ mthp-swapout -s >> >> >> > Iteration 1: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 2: swpout inc: 227, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 3: swpout inc: 222, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 4: swpout inc: 224, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 5: swpout inc: 230, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 6: swpout inc: 229, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 7: swpout inc: 223, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > Iteration 8: swpout inc: 219, swpout fallback inc: 0, Fallback p= ercentage: 0.00% >> >> >> > ... >> >> >> > Iteration 94: swpout inc: 223, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 95: swpout inc: 212, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 96: swpout inc: 220, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 97: swpout inc: 220, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 98: swpout inc: 216, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 99: swpout inc: 223, swpout fallback inc: 0, Fallback = percentage: 0.00% >> >> >> > Iteration 100: swpout inc: 225, swpout fallback inc: 0, Fallback= percentage: 0.00% >> >> >> >> >> >> Unfortunately, the data is gotten using a special designed test pr= ogram >> >> >> which always swap-in pages with swapped-out size. I don't know wh= ether >> >> >> such workloads exist in reality. Otherwise, you need to wait for = mTHP >> >> > >> >> > The test program is designed to simulate mTHP swap behavior using >> >> > zsmalloc and 64KB buffer. >> >> > If we insist on only designing for existing workloads, then zsmalloc >> >> > using 64KB buffer usage will never be able to run, exactly due the >> >> > kernel has high failure rate allocating swap entries for 64KB. There >> >> > is a bit of a chick and egg problem there, such a usage can not exi= st >> >> > because the kernel can't support it yet. Kernel can't add patches to >> >> > support it because such simulation tests are not "real". >> >> > >> >> > We need to break this cycle to support something new. >> >> > >> >> >> swap-in to be merged firstly, and people reach consensus that we s= hould >> >> >> always swap-in pages with swapped-out size. >> >> > >> >> > We don't have to be always. We can identify the situation that makes >> >> > sense. For the zram/zsmalloc 64K buffer usage case, swap out as the >> >> > same swap in size makes sense. >> >> > I think we have agreement on such zsmalloc 64K usage cases we do wa= nt >> >> > to support. >> >> > >> >> >> >> >> >> Alternately, we can make some design adjustment to make the patchs= et >> >> >> work in current situation (mTHP swap-out, normal page swap-in). >> >> >> >> >> >> - One non-full cluster list for each order (same as current design) >> >> >> >> >> >> - When one swap entry is freed, check whether one "order+1" swap e= ntry >> >> >> becomes free, if so, move the cluster to "order+1" non-full clus= ter >> >> >> list. >> >> > >> >> > In the intended zsmalloc usage case, there is no order+1 swap entry >> >> > request. >> >> >> >> This my main concern about this series. Only the Android use cases a= re >> >> considered. The general use cases are just ignored. Is it hard to >> >> consider or test a normal swap partition on your development machine? >> > >> > Please see the V4 cover letter. The V4 already has the SSD, zram and >> > HDD stress testing. >> > Of course I want to make sure the allocator works well with Barry's >> > mthp test case as well. >> > >> >> > Moving the cluster to "order+1" will make less cluster available fo= r "order". >> >> > For that usage case it is negative gain. >> >> >> >> The "order+1" cluster can be used to allocate "order" cluster when >> >> existing "order" cluster is used up. >> >> >> >> And in this way, we can protect clusters with more free spaces so that >> >> they may become free. >> >> >> >> >> - When allocate swap entry with "order", get cluster from free, "o= rder", >> >> >> "order+1", ... non-full cluster list. If all are empty, fallbac= k to >> >> > >> >> > I don't see that it is useful for the zsmalloc 64K buffer usage cas= e. >> >> > There will be order 0 and order 4 and nothing else. >> >> > >> >> > How about let's keep it simple for now. If we identify some workload >> >> > this algorithm can help. We can do that as a follow up step. >> >> >> >> The simple design isn't flexible enough for your workloads too. For >> >> example, >> >> >> >> - Initially, almost only order-0 pages are swapped out, most non-full >> >> clusters are order-0. >> >> >> >> - Later, quite some order-0 swap entries are freed so that there are >> >> quite some order-4 swap entries available. >> >> >> >> - Order-4 pages need to be swapped out, but no enough order-4 non-full >> >> clusters available. >> >> >> >> So, we need a way to migrate non-full clusters among orders to adjust= to >> >> the situations automatically. >> > >> > Depends on how lucky it is to form the order-4 cluster naturally. The >> > odds of forming the order-4 cluster naturally in random swap >> > allocation/ free case is very low. I have the number in my other email >> > thread. >> > Anyway, if we convince this payout for the complexity it introduces, >> > we can do that as follow up steps. Try to keep things simple at first >> > for the review benefit. >> > >> >> >> >> >> order 0. >> >> >> >> >> >> Do you think that this works? >> >> >> >> >> >> > Reported-by: Barry Song <21cnbao@gmail.com> >> >> >> > Signed-off-by: Chris Li >> >> >> > --- >> >> >> > Changes in v3: >> >> >> > - Using V1 as base. >> >> >> > - Rename "next" to "list" for the list field, suggested by Ying. >> >> >> > - Update comment for the locking rules for cluster fields and li= st, >> >> >> > suggested by Ying. >> >> >> > - Allocate from the nonfull list before attempting free list, su= ggested >> >> >> > by Kairui. >> >> >> >> >> >> Haven't looked into this. It appears that this breaks the original >> >> >> discard behavior which helps performance of some SSD, please refer= to >> >> > >> >> > Can you clarify by "discard" you mean SSD discard command or just t= he >> >> > way swap allocator recycles free clusters? >> >> >> >> The SSD discard command, like in the following URL, >> >> >> >> https://en.wikipedia.org/wiki/Trim_(computing) >> > >> > Thanks. I know what an SSD discard command is. Want to understand why >> > that behavior is preferred. >> > >> > So the reasoning to prefer a new free block rather than a recent >> > particle free cluster is to let the previous written cluster have a >> > higher chance to issue the discard command? >> > >> > This preferred new block behavior is actually not friendly to SSD from >> > a wearing point of view. >> > Take this example: >> > Let say the data need to allocate and free from swap. At any given >> > time the swap usage is 1G. The swap SSD drive is 16G. >> > Let say the allocation and free are at random 4K page locations. There >> > is totally 64G swap data needed to write to swap, but at any given >> > time there is only 1G data occupite on swapfile. >> > >> > a) If you always prefer new free blocks. Then the swap data will >> > eventually write at all 16G drives then random write to full 16G. >> > Chance of forming a free cluster so a discard command can be issued is >> > very low. (15/16)**512 =3D 4.4E-15. From SSD point of view, it does not >> > know most of the data written to 16G drive is not used. When a page is >> > free on a swapfile, SSD drive doesn't know about it. It sees 4K random >> > writes to all 16G of the drive, total 64G data written. >> > >> > b) If you always prefer a non full cluster first over a new cluster. >> > The 64G data will concentrate random writing to the first 1G of drive >> > location. Total 64G data written. >> > >> > I consider b) are more friendly to SSD than a). Because concentrate >> > the write into the first 1G location. The SSD can know the data >> > overwritten in those 1G has internally obsolete, so it can internally >> > GC the those overwritten data without a discard command. Where a) >> > random 4K writes to the whole drive without much discard at all. Full >> > SSD doing random writes is a bad combination from a wearing point of >> > view. >> > >> > Just my 2 cents. Anyway I revert the V4 to use free cluster before >> > nonfull cluster just to behave the same as previously. >> > >> >> >> commit 2a8f94493432 ("swap: change block allocation algorithm for = SSD"). >> >> > >> >> > I did read that change log. Help me understand in more detail which >> >> > discard behavior you have in mind. A lot of low end micro SD cards >> >> > have proper FTL wear leveling now, ssd even better on that. >> >> >> >> It's not FTL, it's discard/trim for SSD as above. >> > >> > Thanks for the clarification. >> > >> >> >> >> >> And as pointed out by Ryan, this may reduce the opportunity of the >> >> >> sequential block device writing during swap-out, which may hurt >> >> >> performance of SSD too. >> >> > >> >> > Only at the initial phase. If the swap IO continues, after the first >> >> > pass fills up the swap file, the write will be random on the swapfi= le >> >> > anyway. Because the swapfile only issues 2M discards commands when = all >> >> > 512 4K pages are free. The discarded area will be much smaller than >> >> > the free area on swapfile. That combined with the random write page= on >> >> > the whole swap file. It might produce a worse internal write >> >> > amplification for SSD, compared to only writing a subset of the >> >> > swapfile area. I would love to hear from someone who understands SSD >> >> > internals to confirm or deny my theory. >> >> >> >> It depends on workloads. Some workloads will have more severe >> >> fragmentation than others. For example, on quite some machines, the >> >> swap devices will be far from being full to avoid possible OOM. >> > >> > I suspect most of the SSD swap on client devices nowadays are only as >> > backup just in case it needs to be swapped. >> > There is not much SSD swap IO during normal use. The zram and zswap >> > are more actively used in the data center and Android phone case, from >> > swap IO ops point of view. >> >> I use a Linux laptop with 16GB DRAM for work. And I found that the swap >> space are almost always used. > > Just curious how many swap OPS per second on average? I suspect it > will be a very low number. It depends on workloads. I have run some LLM pruning experiment algorithm on the machine. The swap IOPS is high for that. [snip] -- Best Regards, Huang, Ying