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 X-Spam-Level: X-Spam-Status: No, score=-16.8 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,URIBL_BLOCKED,USER_AGENT_GIT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 819CAC433E0 for ; Thu, 21 Jan 2021 17:18:10 +0000 (UTC) Received: from pdx1-mailman02.dreamhost.com (pdx1-mailman02.dreamhost.com [64.90.62.194]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mail.kernel.org (Postfix) with ESMTPS id 3964D23A5A for ; Thu, 21 Jan 2021 17:18:10 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 3964D23A5A Authentication-Results: mail.kernel.org; dmarc=none (p=none dis=none) header.from=infradead.org Authentication-Results: mail.kernel.org; spf=none smtp.mailfrom=lustre-devel-bounces@lists.lustre.org Received: from pdx1-mailman02.dreamhost.com (localhost [IPv6:::1]) by pdx1-mailman02.dreamhost.com (Postfix) with ESMTP id 2844721FDDD; Thu, 21 Jan 2021 09:17:45 -0800 (PST) Received: from smtp4.ccs.ornl.gov (smtp4.ccs.ornl.gov [160.91.203.40]) by pdx1-mailman02.dreamhost.com (Postfix) with ESMTP id C704521FBB4 for ; Thu, 21 Jan 2021 09:17:10 -0800 (PST) Received: from star.ccs.ornl.gov (star.ccs.ornl.gov [160.91.202.134]) by smtp4.ccs.ornl.gov (Postfix) with ESMTP id 4FDDB1008061; Thu, 21 Jan 2021 12:17:05 -0500 (EST) Received: by star.ccs.ornl.gov (Postfix, from userid 2004) id 4E6571B49C; Thu, 21 Jan 2021 12:17:05 -0500 (EST) From: James Simmons To: Andreas Dilger , Oleg Drokin , NeilBrown Date: Thu, 21 Jan 2021 12:16:40 -0500 Message-Id: <1611249422-556-18-git-send-email-jsimmons@infradead.org> X-Mailer: git-send-email 1.8.3.1 In-Reply-To: <1611249422-556-1-git-send-email-jsimmons@infradead.org> References: <1611249422-556-1-git-send-email-jsimmons@infradead.org> Subject: [lustre-devel] [PATCH 17/39] lustre: llite: try to improve mmap performance X-BeenThere: lustre-devel@lists.lustre.org X-Mailman-Version: 2.1.23 Precedence: list List-Id: "For discussing Lustre software development." List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: Wang Shilong , Lustre Development List MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Errors-To: lustre-devel-bounces@lists.lustre.org Sender: "lustre-devel" From: Wang Shilong We have observed slow mmap read performances for some applications. The problem is if access pattern is neither sequential nor stride, but could be still adjacent in a small range and then seek a random position. So the pattern could be something like this: [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data] Every time an application reads mmap data, it may not only read a single 4KB page, but aslo a cluster of nearby pages in a range(e.g. 1MB) of the first page after a cache miss. The readahead engine is modified to track the range size of a cluster of mmap reads, so that after a seek and/or cache miss, the range size is used to efficiently prefetch multiple pages in a single RPC rather than many small RPCs. Benchmark: fio --name=randread --directory=/ai400/fio --rw=randread --ioengine=mmap --bs=128K --numjobs=32 --filesize=200G --filename=randread --time_based --status-interval=10s --runtime=30s --allow_file_create=1 --group_reporting --disable_lat=1 --disable_clat=1 --disable_slat=1 --disk_util=0 --aux-path=/tmp --randrepeat=0 --unique_filename=0 --fallocate=0 | master | patched | speedup | ---------------+-----------+------------+-----------+ page_fault_avg | 512usec | 52usec | 9.75x page_fault_max | 37698usec| 6543usec| 5.76x WC-bug-id: https://jira.whamcloud.com/browse/LU-13669 Lustre-commit: 0c5ad4b6df5bf3 ("LU-13669 llite: try to improve mmap performance") Signed-off-by: Wang Shilong Reviewed-on: https://review.whamcloud.com/38916 Reviewed-by: Andreas Dilger Reviewed-by: Yingjin Qian Reviewed-by: Oleg Drokin Signed-off-by: James Simmons --- fs/lustre/llite/llite_internal.h | 18 +++++ fs/lustre/llite/llite_lib.c | 1 + fs/lustre/llite/lproc_llite.c | 47 +++++++++++++ fs/lustre/llite/rw.c | 142 +++++++++++++++++++++++++++++++++++---- 4 files changed, 196 insertions(+), 12 deletions(-) diff --git a/fs/lustre/llite/llite_internal.h b/fs/lustre/llite/llite_internal.h index bad974f..797dfea 100644 --- a/fs/lustre/llite/llite_internal.h +++ b/fs/lustre/llite/llite_internal.h @@ -482,6 +482,12 @@ static inline struct pcc_inode *ll_i2pcci(struct inode *inode) /* default read-ahead full files smaller than limit on the second read */ #define SBI_DEFAULT_READ_AHEAD_WHOLE_MAX MiB_TO_PAGES(2UL) +/* default range pages */ +#define SBI_DEFAULT_RA_RANGE_PAGES MiB_TO_PAGES(1ULL) + +/* Min range pages */ +#define RA_MIN_MMAP_RANGE_PAGES 16UL + enum ra_stat { RA_STAT_HIT = 0, RA_STAT_MISS, @@ -498,6 +504,7 @@ enum ra_stat { RA_STAT_FAILED_REACH_END, RA_STAT_ASYNC, RA_STAT_FAILED_FAST_READ, + RA_STAT_MMAP_RANGE_READ, _NR_RA_STAT, }; @@ -505,6 +512,7 @@ struct ll_ra_info { atomic_t ra_cur_pages; unsigned long ra_max_pages; unsigned long ra_max_pages_per_file; + unsigned long ra_range_pages; unsigned long ra_max_read_ahead_whole_pages; struct workqueue_struct *ll_readahead_wq; /* @@ -790,6 +798,16 @@ struct ll_readahead_state { */ pgoff_t ras_window_start_idx; pgoff_t ras_window_pages; + + /* Page index where min range read starts */ + pgoff_t ras_range_min_start_idx; + /* Page index where mmap range read ends */ + pgoff_t ras_range_max_end_idx; + /* number of mmap pages where last time detected */ + pgoff_t ras_last_range_pages; + /* number of mmap range requests */ + pgoff_t ras_range_requests; + /* * Optimal RPC size in pages. * It decides how many pages will be sent for each read-ahead. diff --git a/fs/lustre/llite/llite_lib.c b/fs/lustre/llite/llite_lib.c index 34bd661..c560492 100644 --- a/fs/lustre/llite/llite_lib.c +++ b/fs/lustre/llite/llite_lib.c @@ -130,6 +130,7 @@ static struct ll_sb_info *ll_init_sbi(void) SBI_DEFAULT_READ_AHEAD_PER_FILE_MAX); sbi->ll_ra_info.ra_async_pages_per_file_threshold = sbi->ll_ra_info.ra_max_pages_per_file; + sbi->ll_ra_info.ra_range_pages = SBI_DEFAULT_RA_RANGE_PAGES; sbi->ll_ra_info.ra_max_read_ahead_whole_pages = -1; atomic_set(&sbi->ll_ra_info.ra_async_inflight, 0); diff --git a/fs/lustre/llite/lproc_llite.c b/fs/lustre/llite/lproc_llite.c index 9b1c392..5d1e2f4 100644 --- a/fs/lustre/llite/lproc_llite.c +++ b/fs/lustre/llite/lproc_llite.c @@ -1173,6 +1173,51 @@ static ssize_t read_ahead_async_file_threshold_mb_show(struct kobject *kobj, } LUSTRE_RW_ATTR(read_ahead_async_file_threshold_mb); +static ssize_t read_ahead_range_kb_show(struct kobject *kobj, + struct attribute *attr, char *buf) +{ + struct ll_sb_info *sbi = container_of(kobj, struct ll_sb_info, + ll_kset.kobj); + + return scnprintf(buf, PAGE_SIZE, "%lu\n", + sbi->ll_ra_info.ra_range_pages << (PAGE_SHIFT - 10)); +} + +static ssize_t +read_ahead_range_kb_store(struct kobject *kobj, + struct attribute *attr, + const char *buffer, size_t count) +{ + struct ll_sb_info *sbi = container_of(kobj, struct ll_sb_info, + ll_kset.kobj); + unsigned long pages_number; + unsigned long max_ra_per_file; + u64 val; + int rc; + + rc = sysfs_memparse(buffer, count, &val, "KiB"); + if (rc < 0) + return rc; + + pages_number = val >> PAGE_SHIFT; + /* Disable mmap range read */ + if (pages_number == 0) + goto out; + + max_ra_per_file = sbi->ll_ra_info.ra_max_pages_per_file; + if (pages_number > max_ra_per_file || + pages_number < RA_MIN_MMAP_RANGE_PAGES) + return -ERANGE; + +out: + spin_lock(&sbi->ll_lock); + sbi->ll_ra_info.ra_range_pages = pages_number; + spin_unlock(&sbi->ll_lock); + + return count; +} +LUSTRE_RW_ATTR(read_ahead_range_kb); + static ssize_t fast_read_show(struct kobject *kobj, struct attribute *attr, char *buf) @@ -1506,6 +1551,7 @@ struct ldebugfs_vars lprocfs_llite_obd_vars[] = { &lustre_attr_max_read_ahead_mb.attr, &lustre_attr_max_read_ahead_per_file_mb.attr, &lustre_attr_max_read_ahead_whole_mb.attr, + &lustre_attr_read_ahead_range_kb.attr, &lustre_attr_checksums.attr, &lustre_attr_checksum_pages.attr, &lustre_attr_stats_track_pid.attr, @@ -1622,6 +1668,7 @@ void ll_stats_ops_tally(struct ll_sb_info *sbi, int op, long count) [RA_STAT_FAILED_REACH_END] = "failed to reach end", [RA_STAT_ASYNC] = "async readahead", [RA_STAT_FAILED_FAST_READ] = "failed to fast read", + [RA_STAT_MMAP_RANGE_READ] = "mmap range read", }; int ll_debugfs_register_super(struct super_block *sb, const char *name) diff --git a/fs/lustre/llite/rw.c b/fs/lustre/llite/rw.c index da4a26d..096e015 100644 --- a/fs/lustre/llite/rw.c +++ b/fs/lustre/llite/rw.c @@ -388,7 +388,7 @@ static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria) static unsigned long ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io, struct cl_page_list *queue, struct ll_readahead_state *ras, - struct ra_io_arg *ria, pgoff_t *ra_end) + struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index) { struct cl_read_ahead ra = { 0 }; pgoff_t page_idx; @@ -402,6 +402,8 @@ static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria) for (page_idx = ria->ria_start_idx; page_idx <= ria->ria_end_idx && ria->ria_reserved > 0; page_idx++) { + if (skip_index && page_idx == skip_index) + continue; if (ras_inside_ra_window(page_idx, ria)) { if (!ra.cra_end_idx || ra.cra_end_idx < page_idx) { pgoff_t end_idx; @@ -447,10 +449,12 @@ static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria) if (ras->ras_rpc_pages != ra.cra_rpc_pages && ra.cra_rpc_pages > 0) ras->ras_rpc_pages = ra.cra_rpc_pages; - /* trim it to align with optimal RPC size */ - end_idx = ras_align(ras, ria->ria_end_idx + 1); - if (end_idx > 0 && !ria->ria_eof) - ria->ria_end_idx = end_idx - 1; + if (!skip_index) { + /* trim it to align with optimal RPC size */ + end_idx = ras_align(ras, ria->ria_end_idx + 1); + if (end_idx > 0 && !ria->ria_eof) + ria->ria_end_idx = end_idx - 1; + } if (ria->ria_end_idx < ria->ria_end_idx_min) ria->ria_end_idx = ria->ria_end_idx_min; } @@ -650,7 +654,7 @@ static void ll_readahead_handle_work(struct work_struct *wq) cl_2queue_init(queue); rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria, - &ra_end_idx); + &ra_end_idx, 0); if (ria->ria_reserved != 0) ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved); if (queue->c2_qin.pl_nr > 0) { @@ -688,7 +692,7 @@ static void ll_readahead_handle_work(struct work_struct *wq) static int ll_readahead(const struct lu_env *env, struct cl_io *io, struct cl_page_list *queue, struct ll_readahead_state *ras, bool hit, - struct file *file) + struct file *file, pgoff_t skip_index) { struct vvp_io *vio = vvp_env_io(env); struct ll_thread_info *lti = ll_env_info(env); @@ -731,6 +735,9 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io, if (ras->ras_window_pages > 0) end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1; + if (skip_index) + end_idx = start_idx + ras->ras_window_pages - 1; + /* Enlarge the RA window to encompass the full read */ if (vio->vui_ra_valid && end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1) @@ -783,6 +790,10 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io, ria->ria_start_idx; } + /* don't over reserved for mmap range read */ + if (skip_index) + pages_min = 0; + ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages, pages_min); if (ria->ria_reserved < pages) @@ -793,8 +804,8 @@ static int ll_readahead(const struct lu_env *env, struct cl_io *io, atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages), ll_i2sbi(inode)->ll_ra_info.ra_max_pages); - ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx); - + ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx, + skip_index); if (ria->ria_reserved) ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved); @@ -890,6 +901,10 @@ void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) ras_reset(ras, 0); ras->ras_last_read_end_bytes = 0; ras->ras_requests = 0; + ras->ras_range_min_start_idx = 0; + ras->ras_range_max_end_idx = 0; + ras->ras_range_requests = 0; + ras->ras_last_range_pages = 0; } /* @@ -1033,6 +1048,73 @@ static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos) 8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT); } +static inline bool is_loose_mmap_read(struct ll_sb_info *sbi, + struct ll_readahead_state *ras, + unsigned long pos) +{ + unsigned long range_pages = sbi->ll_ra_info.ra_range_pages; + + return pos_in_window(pos, ras->ras_last_read_end_bytes, + range_pages << PAGE_SHIFT, + range_pages << PAGE_SHIFT); +} + +/** + * We have observed slow mmap read performances for some + * applications. The problem is if access pattern is neither + * sequential nor stride, but could be still adjacent in a + * small range and then seek a random position. + * + * So the pattern could be something like this: + * + * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data] + * + * + * Every time an application reads mmap data, it may not only + * read a single 4KB page, but aslo a cluster of nearby pages in + * a range(e.g. 1MB) of the first page after a cache miss. + * + * The readahead engine is modified to track the range size of + * a cluster of mmap reads, so that after a seek and/or cache miss, + * the range size is used to efficiently prefetch multiple pages + * in a single RPC rather than many small RPCs. + */ +static void ras_detect_cluster_range(struct ll_readahead_state *ras, + struct ll_sb_info *sbi, + unsigned long pos, unsigned long count) +{ + pgoff_t last_pages, pages; + pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT; + + last_pages = ras->ras_range_max_end_idx - + ras->ras_range_min_start_idx + 1; + /* First time come here */ + if (!ras->ras_range_max_end_idx) + goto out; + + /* Random or Stride read */ + if (!is_loose_mmap_read(sbi, ras, pos)) + goto out; + + ras->ras_range_requests++; + if (ras->ras_range_max_end_idx < end_idx) + ras->ras_range_max_end_idx = end_idx; + + if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT)) + ras->ras_range_min_start_idx = pos >> PAGE_SHIFT; + + /* Out of range, consider it as random or stride */ + pages = ras->ras_range_max_end_idx - + ras->ras_range_min_start_idx + 1; + if (pages <= sbi->ll_ra_info.ra_range_pages) + return; +out: + ras->ras_last_range_pages = last_pages; + ras->ras_range_requests = 0; + ras->ras_range_min_start_idx = pos >> PAGE_SHIFT; + ras->ras_range_max_end_idx = end_idx; +} + static void ras_detect_read_pattern(struct ll_readahead_state *ras, struct ll_sb_info *sbi, loff_t pos, size_t count, bool mmap) @@ -1080,9 +1162,13 @@ static void ras_detect_read_pattern(struct ll_readahead_state *ras, ras->ras_consecutive_bytes += count; if (mmap) { + unsigned long ra_range_pages = + max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES, + sbi->ll_ra_info.ra_range_pages); pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT; - if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect) + if ((idx >= ra_range_pages && + idx % ra_range_pages == 0) || stride_detect) ras->ras_need_increase_window = true; } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) { ras->ras_need_increase_window = true; @@ -1190,10 +1276,36 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode, if (ras->ras_no_miss_check) goto out_unlock; - if (flags & LL_RAS_MMAP) + if (flags & LL_RAS_MMAP) { + unsigned long ra_pages; + + ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT, + PAGE_SIZE); ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT, PAGE_SIZE, true); + /* we did not detect anything but we could prefetch */ + if (!ras->ras_need_increase_window && + ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages && + ras->ras_range_requests >= 2) { + if (!hit) { + ra_pages = max_t(unsigned long, + RA_MIN_MMAP_RANGE_PAGES, + ras->ras_last_range_pages); + if (index < ra_pages / 2) + index = 0; + else + index -= ra_pages / 2; + ras->ras_window_pages = ra_pages; + ll_ra_stats_inc_sbi(sbi, + RA_STAT_MMAP_RANGE_READ); + } else { + ras->ras_window_pages = 0; + } + goto skip; + } + } + if (!hit && ras->ras_window_pages && index < ras->ras_next_readahead_idx && pos_in_window(index, ras->ras_window_start_idx, 0, @@ -1231,6 +1343,8 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode, goto out_unlock; } } + +skip: ras_set_start(ras, index); if (stride_io_mode(ras)) { @@ -1500,8 +1614,12 @@ int ll_io_read_page(const struct lu_env *env, struct cl_io *io, io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count - 1); if (ll_readahead_enabled(sbi) && ras) { + pgoff_t skip_index = 0; + + if (ras->ras_next_readahead_idx < vvp_index(vpg)) + skip_index = vvp_index(vpg); rc2 = ll_readahead(env, io, &queue->c2_qin, ras, - uptodate, file); + uptodate, file, skip_index); CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n", PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg)); } else if (vvp_index(vpg) == io_start_index && -- 1.8.3.1 _______________________________________________ lustre-devel mailing list lustre-devel@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org