Project

General

Profile

Tuning for All Flash Deployments » History » Version 9

Patrick McGarry, 01/14/2017 08:34 PM

1 1 Patrick McGarry
h1. Tuning for All Flash Deployments
2 1 Patrick McGarry
3 1 Patrick McGarry
Ceph Tuning and Best Practices for All Flash Intel® Xeon® Servers
4 1 Patrick McGarry
Last updated:  January 2017
5 1 Patrick McGarry
6 1 Patrick McGarry
7 1 Patrick McGarry
8 1 Patrick McGarry
9 1 Patrick McGarry
h2. Table of Contents
10 9 Patrick McGarry
11 9 Patrick McGarry
+[[Tuning_for_All_Flash_Deployments#Introduction|Introduction]]+
12 9 Patrick McGarry
+[[Tuning_for_All_Flash_Deployments#Ceph-Storage-Hardware-Guidelines|Ceph Storage Hardware Guidelines]]+
13 9 Patrick McGarry
+[[Tuning_for_All_Flash_Deployments#Intel-Tuning-and-Optimization-Recommendations-for-Ceph|Intel Tuning and Optimization Recommendations for Ceph]]+
14 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#Server-Tuning|Server Tuning]]
15 8 Patrick McGarry
> > [[Tuning_for_All_Flash_Deployments#Ceph-Client-Configuration|Ceph Client Configuration]]
16 8 Patrick McGarry
> > [[Tuning_for_All_Flash_Deployments#Ceph-Storage-Node-NUMA-Tuning|Ceph Storage Node NUMA Tuning]]
17 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#Memory-Tuning|Memory Tuning]]
18 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#NVMe-SSD-partitioning|NVMe SSD partitioning]]
19 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#OS-Tuning|OS Tuning]] (must be done on all Ceph nodes)
20 8 Patrick McGarry
> > [[Tuning_for_All_Flash_Deployments#Kernel-Tuning|Kernel Tuning]]
21 8 Patrick McGarry
> > [[Tuning_for_All_Flash_Deployments#Filesystem-considerations|Filesystem considerations]]
22 8 Patrick McGarry
> > [[Tuning_for_All_Flash_Deployments#Disk-read-ahead|Disk read ahead]]
23 8 Patrick McGarry
> > [[Tuning_for_All_Flash_Deployments#OSD-RADOS|OSD: RADOS]]
24 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#RBD-Tuning|RBD Tuning]]
25 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#RGW-Rados-Gateway-Tuning|RGW: Rados Gateway Tuning]]
26 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#Erasure-Coding-Tuning|Erasure Coding Tuning]]
27 8 Patrick McGarry
[[Tuning_for_All_Flash_Deployments#Appendix|Appendix]]
28 8 Patrick McGarry
[[Tuning_for_All_Flash_Deployments#Sample-Ceph-conf|Sample Ceph.conf]]
29 8 Patrick McGarry
[[Tuning_for_All_Flash_Deployments#Sample-sysctl-conf|Sample sysctl.conf]]
30 8 Patrick McGarry
[[Tuning_for_All_Flash_Deployments#All-NVMe-Ceph-Cluster-Tuning-for-MySQL-workload|All-NVMe Ceph Cluster Tuning for MySQL workload]]
31 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#Ceph-conf|Ceph.conf]]
32 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#CBT-YAML|CBT YAML]]
33 8 Patrick McGarry
> [[Tuning_for_All_Flash_Deployments#MySQL-configuration-file|MySQL configuration file]] (my.cnf)
34 8 Patrick McGarry
[[Tuning_for_All_Flash_Deployments#Sample-Ceph-Vendor-Solutions|Sample Ceph Vendor Solutions]]+
35 1 Patrick McGarry
36 1 Patrick McGarry
37 1 Patrick McGarry
38 1 Patrick McGarry
39 1 Patrick McGarry
h3. Introduction 
40 1 Patrick McGarry
41 1 Patrick McGarry
Ceph is a scalable, open source, software-defined storage offering that runs on commodity hardware. Ceph has been developed from the ground up to deliver object, block, and file system storage in a single software platform that is self-managing, self-healing and has no single point of failure. Because of its highly scalable, software defined storage architecture, can be a powerful storage solution to consider. 
42 8 Patrick McGarry
This document covers Ceph tuning guidelines specifically for all flash deployments based on extensive testing by Intel with a variety of system, operating system and Ceph optimizations to achieve highest possible performance for servers with Intel® Xeon® processors and Intel® Solid State Drive Data Center (Intel® SSD DC) Series. Details of OEM system SKUs and Ceph reference architectures for targeted use-cases can be found on ceph.com web-site. 
43 8 Patrick McGarry
 
44 8 Patrick McGarry
h3. Ceph Storage Hardware Guidelines  
45 1 Patrick McGarry
46 8 Patrick McGarry
* *Standard* configuration is ideally suited for throughput oriented workloads (e.g., analytics,  DVR). Intel® SSD Data Center P3700 series is recommended to achieve best possible performance while balancing the cost.
47 1 Patrick McGarry
48 8 Patrick McGarry
| CPU | Intel® Xeon® CPU E5-2650v4 or higher |
49 8 Patrick McGarry
| Memory | Minimum of 64 GB|
50 8 Patrick McGarry
| NIC | 10GbE |
51 8 Patrick McGarry
| Disks | 1x 1.6TB P3700 + 12 x 4TB HDDs (1:12 ratio) 
52 8 Patrick McGarry
P3700 as Journal and caching |
53 8 Patrick McGarry
| Caching software | Intel Cache Acceleration Software for read caching, option: Intel® Rapid Storage Technology enterprise/MD4.3 |
54 1 Patrick McGarry
55 8 Patrick McGarry
* *TCO optimized* configuration provides best possible performance for performance centric workloads (e.g., database) while achieving the TCO with a mix of SATA SSDs and NVMe SSDs. 
56 8 Patrick McGarry
57 8 Patrick McGarry
| CPU | Intel® Xeon® CPU E5-2690v4 or higher |
58 8 Patrick McGarry
| Memory | 128 GB or higher |
59 8 Patrick McGarry
| NIC | Dual 10GbE 
60 8 Patrick McGarry
| Disks | 1x 800GB P3700 + 4x S3510 1.6TB |
61 8 Patrick McGarry
62 8 Patrick McGarry
* *IOPS optimized* configuration provides best performance for workloads that demand low latency using all NVMe SSD configuration. 
63 8 Patrick McGarry
64 8 Patrick McGarry
| CPU | Intel® Xeon® CPU E5-2699v4 |
65 8 Patrick McGarry
| Memory | 128 GB or higher |
66 8 Patrick McGarry
| NIC | 1x 40GbE, 4x 10GbE |
67 8 Patrick McGarry
| Disks | 4 x P3700 2TB |
68 8 Patrick McGarry
69 8 Patrick McGarry
H1. Intel Tuning and Optimization Recommendations for Ceph
70 8 Patrick McGarry
71 8 Patrick McGarry
H2. Server Tuning
72 8 Patrick McGarry
73 8 Patrick McGarry
H3. Ceph Client Configuration
74 1 Patrick McGarry
In a balanced system configuration both client and storage node configuration need to be optimized to get the best possible cluster performance. Care needs to be taken to ensure Ceph client node server has enough CPU bandwidth to achieve optimum performance. Below graph shows the end to end performance for different client CPU configurations for block workload.
75 1 Patrick McGarry
76 8 Patrick McGarry
!1-cpu-cores-client.png!
77 1 Patrick McGarry
Figure 1: Client CPU cores and Ceph cluster impact
78 8 Patrick McGarry
79 8 Patrick McGarry
h3. Ceph Storage Node NUMA Tuning
80 8 Patrick McGarry
81 1 Patrick McGarry
In order to avoid latency, it is important to minimize inter-socket communication between NUMA nodes to service client IO as fast as possible and avoid latency penalty.  Based on extensive set of experiments conducted in Intel, it is recommended to pin Ceph OSD processes on the same CPU socket that has NVMe SSDs, HBAs and NIC devices attached. 
82 8 Patrick McGarry
83 8 Patrick McGarry
!2-numa-mode-config.png! 
84 1 Patrick McGarry
Figure 2: NUMA node configuration and OSD assignment
85 1 Patrick McGarry
86 8 Patrick McGarry
> *_Ceph startup scripts need change with setaffinity=" numactl --membind=0 --cpunodebind=0 "_
87 8 Patrick McGarry
88 1 Patrick McGarry
Below performance data shows best possible cluster throughput and lower latency when Ceph OSDs are partitioned by CPU socket to manage media connected to local CPU socket and network IO not going through QPI link. 
89 1 Patrick McGarry
90 8 Patrick McGarry
!3-numa-node-perf-vs-default-sys-config.png!
91 8 Patrick McGarry
!3.1-table.png!
92 1 Patrick McGarry
Figure 3: NUMA node performance compared to default system configuration
93 8 Patrick McGarry
94 8 Patrick McGarry
h3. Memory Tuning
95 8 Patrick McGarry
96 1 Patrick McGarry
Ceph default packages use tcmalloc. For flash optimized configurations, we found jemalloc providing best possible performance without performance degradation over time. Ceph supports jemalloc for the hammer release and later releases but you need to build with jemalloc option enabled.
97 8 Patrick McGarry
98 1 Patrick McGarry
Below graph in figure 4 shows how thread cache size impacts throughput. By tuning thread cache size, performance is comparable between TCMalloc and JEMalloc. However as shown in Figure 5 and Figure 6, TCMalloc performance degrades over time unlike JEMalloc. 
99 1 Patrick McGarry
100 8 Patrick McGarry
!4-thread-cache-size-impact-over-perf.png!
101 1 Patrick McGarry
Figure 4: Thread cache size impact over performance
102 1 Patrick McGarry
103 8 Patrick McGarry
!5.0-tcmalloc-over-time.png!
104 8 Patrick McGarry
!5.1-tcmalloc-over-time.png!
105 1 Patrick McGarry
Figure 5: TCMalloc performance in a running cluster over time
106 1 Patrick McGarry
107 1 Patrick McGarry
108 8 Patrick McGarry
!6.0-jemalloc-over-time.png!
109 8 Patrick McGarry
!6.1-jemalloc-over-time.png!
110 1 Patrick McGarry
Figure 6: JEMalloc performance in a running cluster over time
111 1 Patrick McGarry
112 8 Patrick McGarry
h3. NVMe SSD partitioning
113 1 Patrick McGarry
It is not possible to take advantage of NVMe SSD bandwidth with single OSD.  4 is the optimum number of partitions per SSD drive that gives best possible performance. 
114 1 Patrick McGarry
115 1 Patrick McGarry
Figure 7: Ceph OSD latency with different SSD partitions
116 1 Patrick McGarry
117 1 Patrick McGarry
Figure 8: CPU Utilization with different #of SSD partitions 
118 1 Patrick McGarry
119 1 Patrick McGarry
OS Tuning (must be done on all Ceph nodes)
120 1 Patrick McGarry
Kernel Tuning
121 1 Patrick McGarry
1. Modify system control in /etc/sysctl.conf
122 1 Patrick McGarry
# Kernel sysctl configuration file for Red Hat Linux
123 1 Patrick McGarry
#
124 1 Patrick McGarry
# For binary values, 0 is disabled, 1 is enabled.  See sysctl(8) and
125 1 Patrick McGarry
# sysctl.conf(5) for more details.
126 1 Patrick McGarry
127 1 Patrick McGarry
# Controls IP packet forwarding
128 1 Patrick McGarry
net.ipv4.ip_forward = 0
129 1 Patrick McGarry
130 1 Patrick McGarry
# Controls source route verification
131 1 Patrick McGarry
net.ipv4.conf.default.rp_filter = 1
132 1 Patrick McGarry
133 1 Patrick McGarry
# Do not accept source routing
134 1 Patrick McGarry
net.ipv4.conf.default.accept_source_route = 0
135 1 Patrick McGarry
136 1 Patrick McGarry
# Controls the System Request debugging functionality of the kernel
137 1 Patrick McGarry
kernel.sysrq = 0
138 1 Patrick McGarry
139 1 Patrick McGarry
# Controls whether core dumps will append the PID to the core filename.
140 1 Patrick McGarry
# Useful for debugging multi-threaded applications.
141 1 Patrick McGarry
kernel.core_uses_pid = 1
142 1 Patrick McGarry
143 1 Patrick McGarry
# disable TIME_WAIT.. wait ..
144 1 Patrick McGarry
net.ipv4.tcp_tw_recycle = 1
145 1 Patrick McGarry
net.ipv4.tcp_tw_reuse = 1
146 1 Patrick McGarry
147 1 Patrick McGarry
# Controls the use of TCP syncookies
148 1 Patrick McGarry
net.ipv4.tcp_syncookies = 0
149 1 Patrick McGarry
150 1 Patrick McGarry
# double amount of allowed conntrack
151 1 Patrick McGarry
net.netfilter.nf_conntrack_max = 2621440
152 1 Patrick McGarry
net.netfilter.nf_conntrack_tcp_timeout_established = 1800
153 1 Patrick McGarry
154 1 Patrick McGarry
# Disable netfilter on bridges.
155 1 Patrick McGarry
net.bridge.bridge-nf-call-ip6tables = 0
156 1 Patrick McGarry
net.bridge.bridge-nf-call-iptables = 0
157 1 Patrick McGarry
net.bridge.bridge-nf-call-arptables = 0
158 1 Patrick McGarry
159 1 Patrick McGarry
# Controls the maximum size of a message, in bytes
160 1 Patrick McGarry
kernel.msgmnb = 65536
161 1 Patrick McGarry
162 1 Patrick McGarry
# Controls the default maxmimum size of a mesage queue
163 1 Patrick McGarry
kernel.msgmax = 65536
164 1 Patrick McGarry
165 1 Patrick McGarry
# Controls the maximum shared segment size, in bytes
166 1 Patrick McGarry
kernel.shmmax = 68719476736
167 1 Patrick McGarry
168 1 Patrick McGarry
# Controls the maximum number of shared memory segments, in pages
169 1 Patrick McGarry
kernel.shmall = 4294967296
170 1 Patrick McGarry
171 1 Patrick McGarry
2. IP jumbo frames
172 1 Patrick McGarry
If your switch supports jumbo frames, then the larger MTU size is helpful. Our tests showed 9000 MTU improves Sequential Read/Write performance.
173 1 Patrick McGarry
174 1 Patrick McGarry
3. Set the Linux disk scheduler to cfq
175 1 Patrick McGarry
Filesystem considerations
176 1 Patrick McGarry
Ceph is designed to be mostly filesystem agnostic–the only requirement being that the filesystem supports extended attributes (xattrs). Ceph OSDs depend on the Extended Attributes (XATTRs) of the underlying file system for: a) Internal object state b) Snapshot metadata c) RGW Access control Lists etc. Currently XFS is the recommended file system. We recommend using big inode size (default inode size is 256 bytes) when creating the file system:
177 1 Patrick McGarry
mkfs.xfs –i size=2048 /dev/sda1
178 1 Patrick McGarry
Setting the inode size is important, as XFS stores xattr data in the inode. If the metadata is too large to fit in the inode, a new extent is created, which can cause quite a performance problem. Upping the inode size to 2048 bytes provides enough room to write the default metadata, plus a little headroom.
179 1 Patrick McGarry
The following example mount options are recommended when using XFS:
180 1 Patrick McGarry
mount -t xfs -o noatime,nodiratime,nobarrier,logbufs=8 /dev/sda1 /var/lib/Ceph/osd/Ceph-0
181 1 Patrick McGarry
The following are specific recommendations for Intel SSD and Ceph. 
182 1 Patrick McGarry
mkfs.xfs -f -K -i size=2048 -s size=4096 /dev/md0
183 1 Patrick McGarry
/bin/mount -o noatime,nodiratime,nobarrier /dev/md0 /data/mysql
184 1 Patrick McGarry
Disk read ahead
185 1 Patrick McGarry
Read_ahead is the file prefetching technology used in the Linux operating system. It is a system call that loads a file's contents into the page cache. When a file is subsequently accessed, its contents are read from physical memory rather than from disk, which is much faster.
186 1 Patrick McGarry
echo 2048 > /sys/block/${disk}/queue/read_ahead_kb  (default 128)
187 1 Patrick McGarry
188 1 Patrick McGarry
Per disk performance
189 1 Patrick McGarry
128
190 1 Patrick McGarry
512
191 1 Patrick McGarry
%
192 1 Patrick McGarry
Sequential Read(MB/s)
193 1 Patrick McGarry
1232 MB/s
194 1 Patrick McGarry
3251 MB/s
195 1 Patrick McGarry
+163%
196 1 Patrick McGarry
* 6 nodes Ceph cluster, each have 20 OSD (750 GB * 7200 RPM. 2.5’’ HDD)
197 1 Patrick McGarry
198 1 Patrick McGarry
OSD: RADOS
199 1 Patrick McGarry
Tuning have significant performance impact of Ceph storage system, there are hundreds of tuning knobs for swift. We will introduce some of the most important tuning settings.
200 1 Patrick McGarry
1. Large PG/PGP number (since Cuttlefish)
201 1 Patrick McGarry
We find using large PG number per OSD (>200) will improve the performance. Also this will ease the data distribution unbalance issue
202 1 Patrick McGarry
(default to 8)
203 1 Patrick McGarry
ceph osd pool create testpool 8192 8192
204 1 Patrick McGarry
205 1 Patrick McGarry
2. omap data on separate disks (since Giant)
206 1 Patrick McGarry
Mounting omap directory to some separate SSD will improve the random write performance. In our testing we saw a ~20% performance improvement.
207 1 Patrick McGarry
208 1 Patrick McGarry
3. objecter_inflight_ops/objecter_inflight_op_bytes (since Cuttlefish)
209 1 Patrick McGarry
objecter_inflight_ops/objecter_inflight_op_bytes throttles tell objecter to throttle outgoing ops according its budget, objecter is responsible for send requests to OSD. By default tweak this parameter to 10x 
210 1 Patrick McGarry
(default to 1024/1024*1024*100)
211 1 Patrick McGarry
objecter_inflight_ops = 10240
212 1 Patrick McGarry
objecter_inflight_op_bytes = 1048576000
213 1 Patrick McGarry
214 1 Patrick McGarry
4. ms_dispatch_throttle_bytes (since Cuttlefish)
215 1 Patrick McGarry
ms_dispatch_throttle_bytes throttle is to throttle dispatch message size for simple messenger, by default tweak this parameter to 10x. 
216 1 Patrick McGarry
ms_dispatch_throttle_bytes = 1048576000
217 1 Patrick McGarry
218 1 Patrick McGarry
5. journal_queue_max_bytes/journal_queue_max_ops (since Cuttlefish)
219 1 Patrick McGarry
journal_queue_max_bytes/journal_queue_max_op throttles are to throttle inflight ops for journal, 
220 1 Patrick McGarry
If journal does not get enough budget for current op, it will block osd op thread, by default tweak this parameter to 10x.
221 1 Patrick McGarry
journal_queueu_max_ops = 3000
222 1 Patrick McGarry
journal_queue_max_bytes = 1048576000
223 1 Patrick McGarry
224 1 Patrick McGarry
225 1 Patrick McGarry
6. filestore_queue_max_ops/filestore_queue_max_bytes (since Cuttlefish)
226 1 Patrick McGarry
filestore_queue_max_ops/filestore_queue_max_bytes throttle are used to throttle inflight ops for filestore, these throttles are checked before sending ops to journal, so if filestore does not get enough budget for current op, osd op thread will be blocked, by default tweak this parameter to 10x.
227 1 Patrick McGarry
filestore_queue_max_ops=5000
228 1 Patrick McGarry
filestore_queue_max_bytes = 1048576000
229 1 Patrick McGarry
230 1 Patrick McGarry
7. filestore_op_threads controls the number of filesystem operation threads that execute in parallel
231 1 Patrick McGarry
If the storage backend is fast enough and has enough queues to support parallel operations, it’s recommended to increase this parameter, given there is enough CPU head room.
232 1 Patrick McGarry
filestore_op_threads=6
233 1 Patrick McGarry
234 1 Patrick McGarry
8. journal_max_write_entries/journal_max_write_bytes (since Cuttlefish)
235 1 Patrick McGarry
journal_max_write_entries/journal_max_write_bytes throttle are used to throttle ops or bytes for every journal write, tweaking these two parameters maybe helpful for small write, by default tweak these two parameters to 10x
236 1 Patrick McGarry
journal_max_write_entries = 5000
237 1 Patrick McGarry
journal_max_write_bytes = 1048576000
238 1 Patrick McGarry
239 1 Patrick McGarry
       
240 1 Patrick McGarry
9. osd_op_num_threads_per_shard/osd_op_num_shards (since Firefly)
241 1 Patrick McGarry
osd_op_num_shards set number of queues to cache requests , osd_op_num_threads_per_shard is    threads number for each queue,  adjusting these two parameters depends on cluster.
242 1 Patrick McGarry
After several performance tests with different settings, we concluded that default parameters provide best performance.
243 1 Patrick McGarry
244 1 Patrick McGarry
10. filestore_max_sync_interval (since Cuttlefish)
245 1 Patrick McGarry
filestore_max_sync_interval control the interval that sync thread flush data from memory to disk, by default filestore write data to memory and sync thread is responsible for flushing data to disk, then journal entries can be trimmed. Note that large filestore_max_sync_interval can cause performance spike. By default tweak this parameter to 10 seconds
246 1 Patrick McGarry
filestore_max_sync_interval = 10
247 1 Patrick McGarry
248 1 Patrick McGarry
249 1 Patrick McGarry
11. ms_crc_data/ms_crc_header (since Cuttlefish)
250 1 Patrick McGarry
Disable crc computation for simple messenger, this can reduce CPU utilization
251 1 Patrick McGarry
252 1 Patrick McGarry
12. filestore_fd_cache_shards/filestore_fd_cache_size (since Firefly)
253 1 Patrick McGarry
filestore cache is map from objectname to fd, filestore_fd_cache_shards set number of LRU Cache,  filestore_fd_cache_size is cache size, tweak these two parameter maybe reduce lookup time of fd
254 1 Patrick McGarry
255 1 Patrick McGarry
256 1 Patrick McGarry
13. Set debug level to 0 (since Cuttlefish)
257 1 Patrick McGarry
For an all-SSD Ceph cluster, set debug level for sub system to 0 will improve the performance.  
258 1 Patrick McGarry
debug_lockdep = 0/0
259 1 Patrick McGarry
debug_context = 0/0
260 1 Patrick McGarry
debug_crush = 0/0
261 1 Patrick McGarry
debug_buffer = 0/0
262 1 Patrick McGarry
debug_timer = 0/0
263 1 Patrick McGarry
debug_filer = 0/0
264 1 Patrick McGarry
debug_objecter = 0/0
265 1 Patrick McGarry
debug_rados = 0/0
266 1 Patrick McGarry
debug_rbd = 0/0
267 1 Patrick McGarry
debug_journaler = 0/0
268 1 Patrick McGarry
debug_objectcatcher = 0/0
269 1 Patrick McGarry
debug_client = 0/0
270 1 Patrick McGarry
debug_osd = 0/0
271 1 Patrick McGarry
debug_optracker = 0/0
272 1 Patrick McGarry
debug_objclass = 0/0
273 1 Patrick McGarry
debug_filestore = 0/0
274 1 Patrick McGarry
debug_journal = 0/0
275 1 Patrick McGarry
debug_ms = 0/0
276 1 Patrick McGarry
debug_monc = 0/0
277 1 Patrick McGarry
debug_tp = 0/0
278 1 Patrick McGarry
debug_auth = 0/0
279 1 Patrick McGarry
debug_finisher = 0/0
280 1 Patrick McGarry
debug_heartbeatmap = 0/0
281 1 Patrick McGarry
debug_perfcounter = 0/0
282 1 Patrick McGarry
debug_asok = 0/0
283 1 Patrick McGarry
debug_throttle = 0/0
284 1 Patrick McGarry
debug_mon = 0/0
285 1 Patrick McGarry
debug_paxos = 0/0
286 1 Patrick McGarry
debug_rgw = 0/0
287 1 Patrick McGarry
288 1 Patrick McGarry
289 1 Patrick McGarry
RBD Tuning
290 1 Patrick McGarry
To help achieve low latency on their RBD layer, we suggest the following, in addition to the CERN tuning referenced in ceph.com. 
291 1 Patrick McGarry
1) echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor /dev/null 
292 1 Patrick McGarry
2) start each ceph-osd in dedicated cgroup with dedicated cpu cores (which should be free from any other load, even the kernel one like network interrupts)
293 1 Patrick McGarry
3) increase “filestore_omap_header_cache_size” • “filestore_fd_cache_size” , for better caching (16MB for each 500GB of storage)
294 1 Patrick McGarry
For disk entry in libvirt  put address to all three ceph monitors.
295 1 Patrick McGarry
296 1 Patrick McGarry
RGW: Rados Gateway Tuning
297 1 Patrick McGarry
1. Disable usage/access log (since Cuttlefish)
298 1 Patrick McGarry
rgw enable ops log = false
299 1 Patrick McGarry
rgw enable usage log = false
300 1 Patrick McGarry
log file = /dev/null
301 1 Patrick McGarry
We find disabling usage/access log improves the performance.
302 1 Patrick McGarry
2. Using large cache size (since Cuttlefish)
303 1 Patrick McGarry
rgw cache enabled = true
304 1 Patrick McGarry
rgw cache lru size = 100000
305 1 Patrick McGarry
Caching the hot objects improves the GET performance.
306 1 Patrick McGarry
3. Using larger PG split/merge value.  (since Firefly)
307 1 Patrick McGarry
filestore_merge_threshold = 500
308 1 Patrick McGarry
filestore_split_multiple = 100
309 1 Patrick McGarry
We find PG split/merge will introduce a big overhead. Using a large value would postpone the split/merge behavior. This will help the case where lots of small files are stored in the cluster.
310 1 Patrick McGarry
4. Using load balancer with multiple RGW instances (since Cuttlefish)
311 1 Patrick McGarry
312 1 Patrick McGarry
We’ve found that the RGW has some scalability issues at present. With a single RGW instance the performance is poor. Running multiple RGW instances with a load balancer (e.g., Haproxy) will greatly improve the throughput.
313 1 Patrick McGarry
5. Increase the number of Rados handlers (since Hammer)
314 1 Patrick McGarry
Since Hammer it’s able to using multiple number of Rados handlers per RGW instances. Increasing this value should improve the performance.
315 1 Patrick McGarry
6. Using Civetweb frontend (since Giant)
316 1 Patrick McGarry
Before Giant, Apache + Libfastcgi were the recommended settings. However libfastcgi still use the very old ‘select’ mode, which is not able to handle large amount of concurrent IO in our testing. Using Civetweb frontend would help to improve the stability.
317 1 Patrick McGarry
rgw frontends =civetweb port=80
318 1 Patrick McGarry
319 1 Patrick McGarry
7. Moving bucket index to SSD (since Giant)
320 1 Patrick McGarry
Bucket index updating maybe some bottleneck if there’s millions of objects in one single bucket. We’ve find moving the bucket index to SSD storage will improve the performance.
321 1 Patrick McGarry
322 1 Patrick McGarry
8. Bucket Index Sharding (since Hammer)
323 1 Patrick McGarry
We’ve find the bucket index sharding is a problem if there’s large amount of objects inside one bucket. However the index listing speed may be impacted.
324 1 Patrick McGarry
325 1 Patrick McGarry
Erasure Coding Tuning
326 1 Patrick McGarry
1. Use larger stripe width 
327 1 Patrick McGarry
The default erasure code stripe size (4K) is not optimal, We find using a bigger value (64K) will reduce the CPU% a lot (10%+)
328 1 Patrick McGarry
osd_pool_erasure_code_stripe_width = 65536
329 1 Patrick McGarry
330 1 Patrick McGarry
2. Use mid-sized K
331 1 Patrick McGarry
For the Erasure Code algorithms, we find using some mid-sized K value would bring balanced results between throughput and CPU%. We recommend to use 10+4 or 8+2 mode
332 1 Patrick McGarry
Appendix
333 1 Patrick McGarry
334 1 Patrick McGarry
Sample Ceph.conf 
335 1 Patrick McGarry
[global]
336 1 Patrick McGarry
fsid = 35b08d01-b688-4b9a-947b-bc2e25719370
337 1 Patrick McGarry
mon_initial_members = gw2
338 1 Patrick McGarry
mon_host = 10.10.10.105
339 1 Patrick McGarry
filestore_xattr_use_omap = true
340 1 Patrick McGarry
auth_cluster_required = none
341 1 Patrick McGarry
auth_service_required = none
342 1 Patrick McGarry
auth_client_required = none
343 1 Patrick McGarry
debug_lockdep = 0/0
344 1 Patrick McGarry
debug_context = 0/0
345 1 Patrick McGarry
debug_crush = 0/0
346 1 Patrick McGarry
debug_buffer = 0/0
347 1 Patrick McGarry
debug_timer = 0/0
348 1 Patrick McGarry
debug_filer = 0/0
349 1 Patrick McGarry
debug_objecter = 0/0
350 1 Patrick McGarry
debug_rados = 0/0
351 1 Patrick McGarry
debug_rbd = 0/0
352 1 Patrick McGarry
debug_journaler = 0/0
353 1 Patrick McGarry
debug_objectcatcher = 0/0
354 1 Patrick McGarry
debug_client = 0/0
355 1 Patrick McGarry
debug_osd = 0/0
356 1 Patrick McGarry
debug_optracker = 0/0
357 1 Patrick McGarry
debug_objclass = 0/0
358 1 Patrick McGarry
debug_filestore = 0/0
359 1 Patrick McGarry
debug_journal = 0/0
360 1 Patrick McGarry
debug_ms = 0/0
361 1 Patrick McGarry
debug_monc = 0/0
362 1 Patrick McGarry
debug_tp = 0/0
363 1 Patrick McGarry
debug_auth = 0/0
364 1 Patrick McGarry
debug_finisher = 0/0
365 1 Patrick McGarry
debug_heartbeatmap = 0/0
366 1 Patrick McGarry
debug_perfcounter = 0/0
367 1 Patrick McGarry
debug_asok = 0/0
368 1 Patrick McGarry
debug_throttle = 0/0
369 1 Patrick McGarry
debug_mon = 0/0
370 1 Patrick McGarry
debug_paxos = 0/0
371 1 Patrick McGarry
debug_rgw = 0/0
372 1 Patrick McGarry
[mon]
373 1 Patrick McGarry
mon_pg_warn_max_per_osd=5000
374 1 Patrick McGarry
mon_max_pool_pg_num=106496
375 1 Patrick McGarry
[client]
376 1 Patrick McGarry
rbd cache = false
377 1 Patrick McGarry
[osd]
378 1 Patrick McGarry
osd mkfs type = xfs
379 1 Patrick McGarry
osd mount options xfs = rw,noatime,,nodiratime,inode64,logbsize=256k,delaylog
380 1 Patrick McGarry
osd mkfs options xfs = -f -i size=2048
381 1 Patrick McGarry
filestore_queue_max_ops=5000
382 1 Patrick McGarry
filestore_queue_max_bytes = 1048576000
383 1 Patrick McGarry
filestore_max_sync_interval = 10
384 1 Patrick McGarry
filestore_merge_threshold = 500
385 1 Patrick McGarry
filestore_split_multiple = 100
386 1 Patrick McGarry
osd_op_shard_threads = 8
387 1 Patrick McGarry
journal_max_write_entries = 5000
388 1 Patrick McGarry
journal_max_write_bytes = 1048576000
389 1 Patrick McGarry
journal_queueu_max_ops = 3000
390 1 Patrick McGarry
journal_queue_max_bytes = 1048576000
391 1 Patrick McGarry
ms_dispatch_throttle_bytes = 1048576000
392 1 Patrick McGarry
objecter_inflight_op_bytes = 1048576000
393 1 Patrick McGarry
public_network = 10.10.10.100/24
394 1 Patrick McGarry
cluster_network = 10.10.10.100/24
395 1 Patrick McGarry
396 1 Patrick McGarry
[client.radosgw.gw2-1]
397 1 Patrick McGarry
host = gw2
398 1 Patrick McGarry
keyring = /etc/ceph/ceph.client.radosgw.keyring
399 1 Patrick McGarry
rgw cache enabled = true
400 1 Patrick McGarry
rgw cache lru size = 100000
401 1 Patrick McGarry
rgw socket path = /var/run/ceph/ceph.client.radosgw.gw2-1.fastcgi.sock
402 1 Patrick McGarry
rgw thread pool size = 256
403 1 Patrick McGarry
rgw enable ops log = false
404 1 Patrick McGarry
rgw enable usage log = false
405 1 Patrick McGarry
log file = /dev/null
406 1 Patrick McGarry
rgw frontends =civetweb port=80
407 1 Patrick McGarry
rgw override bucket index max shards = 8
408 1 Patrick McGarry
409 1 Patrick McGarry
Sample sysctl.conf 
410 1 Patrick McGarry
fs.file-max = 6553600
411 1 Patrick McGarry
net.ipv4.ip_local_port_range = 1024 65000
412 1 Patrick McGarry
net.ipv4.tcp_fin_timeout = 20
413 1 Patrick McGarry
net.ipv4.tcp_max_syn_backlog = 819200
414 1 Patrick McGarry
net.ipv4.tcp_keepalive_time = 20
415 1 Patrick McGarry
kernel.msgmni = 2878
416 1 Patrick McGarry
kernel.sem = 256 32000 100 142
417 1 Patrick McGarry
kernel.shmmni = 4096
418 1 Patrick McGarry
net.core.rmem_default = 1048576
419 1 Patrick McGarry
net.core.rmem_max = 1048576
420 1 Patrick McGarry
net.core.wmem_default = 1048576
421 1 Patrick McGarry
net.core.wmem_max = 1048576
422 1 Patrick McGarry
net.core.somaxconn = 40000
423 1 Patrick McGarry
net.core.netdev_max_backlog = 300000
424 1 Patrick McGarry
net.ipv4.tcp_max_tw_buckets = 10000
425 1 Patrick McGarry
426 1 Patrick McGarry
All-NVMe Ceph Cluster Tuning for MySQL workload
427 1 Patrick McGarry
Ceph.conf 
428 1 Patrick McGarry
[global]
429 1 Patrick McGarry
        enable experimental unrecoverable data corrupting features = bluestore rocksdb
430 1 Patrick McGarry
        osd objectstore = bluestore
431 1 Patrick McGarry
        ms_type = async
432 1 Patrick McGarry
        rbd readahead disable after bytes = 0
433 1 Patrick McGarry
        rbd readahead max bytes = 4194304
434 1 Patrick McGarry
        bluestore default buffered read = true
435 1 Patrick McGarry
        auth client required = none
436 1 Patrick McGarry
        auth cluster required = none
437 1 Patrick McGarry
        auth service required = none
438 1 Patrick McGarry
        filestore xattr use omap = true
439 1 Patrick McGarry
        cluster network = 192.168.142.0/24, 192.168.143.0/24
440 1 Patrick McGarry
        private network = 192.168.144.0/24, 192.168.145.0/24
441 1 Patrick McGarry
        log file = /var/log/ceph/$name.log
442 1 Patrick McGarry
        log to syslog = false
443 1 Patrick McGarry
        mon compact on trim = false
444 1 Patrick McGarry
        osd pg bits = 8
445 1 Patrick McGarry
        osd pgp bits = 8
446 1 Patrick McGarry
        mon pg warn max object skew = 100000
447 1 Patrick McGarry
        mon pg warn min per osd = 0
448 1 Patrick McGarry
        mon pg warn max per osd = 32768
449 1 Patrick McGarry
        debug_lockdep = 0/0
450 1 Patrick McGarry
        debug_context = 0/0
451 1 Patrick McGarry
        debug_crush = 0/0
452 1 Patrick McGarry
        debug_buffer = 0/0
453 1 Patrick McGarry
        debug_timer = 0/0
454 1 Patrick McGarry
        debug_filer = 0/0
455 1 Patrick McGarry
        debug_objecter = 0/0
456 1 Patrick McGarry
        debug_rados = 0/0
457 1 Patrick McGarry
        debug_rbd = 0/0
458 1 Patrick McGarry
        debug_ms = 0/0
459 1 Patrick McGarry
        debug_monc = 0/0
460 1 Patrick McGarry
        debug_tp = 0/0
461 1 Patrick McGarry
        debug_auth = 0/0
462 1 Patrick McGarry
        debug_finisher = 0/0
463 1 Patrick McGarry
        debug_heartbeatmap = 0/0
464 1 Patrick McGarry
        debug_perfcounter = 0/0
465 1 Patrick McGarry
        debug_asok = 0/0
466 1 Patrick McGarry
        debug_throttle = 0/0
467 1 Patrick McGarry
        debug_mon = 0/0
468 1 Patrick McGarry
        debug_paxos = 0/0
469 1 Patrick McGarry
        debug_rgw = 0/0
470 1 Patrick McGarry
        perf = true
471 1 Patrick McGarry
        mutex_perf_counter = true
472 1 Patrick McGarry
        throttler_perf_counter = false
473 1 Patrick McGarry
        rbd cache = false
474 1 Patrick McGarry
[mon]
475 1 Patrick McGarry
        mon data =/home/bmpa/tmp_cbt/ceph/mon.$id
476 1 Patrick McGarry
        mon_max_pool_pg_num=166496
477 1 Patrick McGarry
        mon_osd_max_split_count = 10000
478 1 Patrick McGarry
        mon_pg_warn_max_per_osd = 10000
479 1 Patrick McGarry
[mon.a]
480 1 Patrick McGarry
        host = ft02
481 1 Patrick McGarry
        mon addr = 192.168.142.202:6789
482 1 Patrick McGarry
[osd]
483 1 Patrick McGarry
        osd_mount_options_xfs = rw,noatime,inode64,logbsize=256k,delaylog
484 1 Patrick McGarry
        osd_mkfs_options_xfs = -f -i size=2048
485 1 Patrick McGarry
        osd_op_threads = 32
486 1 Patrick McGarry
        filestore_queue_max_ops=5000
487 1 Patrick McGarry
        filestore_queue_committing_max_ops=5000
488 1 Patrick McGarry
        journal_max_write_entries=1000
489 1 Patrick McGarry
        journal_queue_max_ops=3000
490 1 Patrick McGarry
        objecter_inflight_ops=102400
491 1 Patrick McGarry
        filestore_wbthrottle_enable=false
492 1 Patrick McGarry
        filestore_queue_max_bytes=1048576000
493 1 Patrick McGarry
        filestore_queue_committing_max_bytes=1048576000
494 1 Patrick McGarry
        journal_max_write_bytes=1048576000
495 1 Patrick McGarry
        journal_queue_max_bytes=1048576000
496 1 Patrick McGarry
        ms_dispatch_throttle_bytes=1048576000
497 1 Patrick McGarry
        objecter_infilght_op_bytes=1048576000
498 1 Patrick McGarry
        osd_mkfs_type = xfs
499 1 Patrick McGarry
        filestore_max_sync_interval=10
500 1 Patrick McGarry
        osd_client_message_size_cap = 0
501 1 Patrick McGarry
        osd_client_message_cap = 0
502 1 Patrick McGarry
        osd_enable_op_tracker = false
503 1 Patrick McGarry
        filestore_fd_cache_size = 64
504 1 Patrick McGarry
        filestore_fd_cache_shards = 32
505 1 Patrick McGarry
        filestore_op_threads = 6
506 1 Patrick McGarry
507 1 Patrick McGarry
CBT YAML
508 1 Patrick McGarry
cluster:
509 1 Patrick McGarry
  user: "bmpa"
510 1 Patrick McGarry
  head: "ft01"
511 1 Patrick McGarry
  clients: ["ft01", "ft02", "ft03", "ft04", "ft05", "ft06"]
512 1 Patrick McGarry
  osds: ["hswNode01", "hswNode02", "hswNode03", "hswNode04", "hswNode05"]
513 1 Patrick McGarry
  mons:
514 1 Patrick McGarry
   ft02:
515 1 Patrick McGarry
     a: "192.168.142.202:6789"
516 1 Patrick McGarry
osds_per_node: 16
517 1 Patrick McGarry
  fs: xfs
518 1 Patrick McGarry
  mkfs_opts: '-f -i size=2048 -n size=64k'
519 1 Patrick McGarry
  mount_opts: '-o inode64,noatime,logbsize=256k'
520 1 Patrick McGarry
  conf_file: '/home/bmpa/cbt/ceph.conf'
521 1 Patrick McGarry
  use_existing: False
522 1 Patrick McGarry
  newstore_block: True
523 1 Patrick McGarry
  rebuild_every_test: False
524 1 Patrick McGarry
  clusterid: "ceph"
525 1 Patrick McGarry
iterations: 1
526 1 Patrick McGarry
  tmp_dir: "/home/bmpa/tmp_cbt"
527 1 Patrick McGarry
pool_profiles:
528 1 Patrick McGarry
    2rep:
529 1 Patrick McGarry
      pg_size: 8192
530 1 Patrick McGarry
      pgp_size: 8192
531 1 Patrick McGarry
      replication: 2
532 1 Patrick McGarry
benchmarks:
533 1 Patrick McGarry
  librbdfio:
534 1 Patrick McGarry
    time: 300
535 1 Patrick McGarry
    ramp: 300
536 1 Patrick McGarry
    vol_size: 10
537 1 Patrick McGarry
    mode: ['randrw']
538 1 Patrick McGarry
    rwmixread: [0,70,100]
539 1 Patrick McGarry
    op_size: [4096]
540 1 Patrick McGarry
    procs_per_volume: [1]
541 1 Patrick McGarry
    volumes_per_client: [10]
542 1 Patrick McGarry
    use_existing_volumes: False
543 1 Patrick McGarry
    iodepth: [4,8,16,32,64,128]
544 1 Patrick McGarry
    osd_ra: [4096]
545 1 Patrick McGarry
    norandommap: True
546 1 Patrick McGarry
    cmd_path: '/usr/local/bin/fio'
547 1 Patrick McGarry
    pool_profile: '2rep'
548 1 Patrick McGarry
    log_avg_msec: 250
549 1 Patrick McGarry
550 1 Patrick McGarry
MySQL configuration file (my.cnf)
551 1 Patrick McGarry
[client]
552 1 Patrick McGarry
port            = 3306
553 1 Patrick McGarry
socket          = /var/run/mysqld/mysqld.sock
554 1 Patrick McGarry
[mysqld_safe]
555 1 Patrick McGarry
socket          = /var/run/mysqld/mysqld.sock
556 1 Patrick McGarry
nice            = 0
557 1 Patrick McGarry
[mysqld]
558 1 Patrick McGarry
user            = mysql
559 1 Patrick McGarry
pid-file        = /var/run/mysqld/mysqld.pid
560 1 Patrick McGarry
socket          = /var/run/mysqld/mysqld.sock
561 1 Patrick McGarry
port            = 3306
562 1 Patrick McGarry
datadir         = /data
563 1 Patrick McGarry
basedir         = /usr
564 1 Patrick McGarry
tmpdir          = /tmp
565 1 Patrick McGarry
lc-messages-dir = /usr/share/mysql
566 1 Patrick McGarry
skip-external-locking
567 1 Patrick McGarry
bind-address            = 0.0.0.0
568 1 Patrick McGarry
max_allowed_packet      = 16M
569 1 Patrick McGarry
thread_stack            = 192K
570 1 Patrick McGarry
thread_cache_size       = 8
571 1 Patrick McGarry
query_cache_limit       = 1M
572 1 Patrick McGarry
query_cache_size        = 16M
573 1 Patrick McGarry
log_error = /var/log/mysql/error.log
574 1 Patrick McGarry
expire_logs_days        = 10
575 1 Patrick McGarry
max_binlog_size         = 100M
576 1 Patrick McGarry
performance_schema=off
577 1 Patrick McGarry
innodb_buffer_pool_size = 25G
578 1 Patrick McGarry
innodb_flush_method = O_DIRECT
579 1 Patrick McGarry
innodb_log_file_size=4G
580 1 Patrick McGarry
thread_cache_size=16
581 1 Patrick McGarry
innodb_file_per_table
582 1 Patrick McGarry
innodb_checksums = 0
583 1 Patrick McGarry
innodb_flush_log_at_trx_commit = 0
584 1 Patrick McGarry
innodb_write_io_threads = 8
585 1 Patrick McGarry
innodb_page_cleaners= 16
586 1 Patrick McGarry
innodb_read_io_threads = 8
587 1 Patrick McGarry
max_connections = 50000
588 1 Patrick McGarry
[mysqldump]
589 1 Patrick McGarry
quick
590 1 Patrick McGarry
quote-names
591 1 Patrick McGarry
max_allowed_packet      = 16M
592 1 Patrick McGarry
[mysql]
593 1 Patrick McGarry
!includedir /etc/mysql/conf.d/
594 1 Patrick McGarry
595 1 Patrick McGarry
596 1 Patrick McGarry
Sample Ceph Vendor Solutions
597 1 Patrick McGarry
The following are pointers to Ceph solutions, but this list is not comprehensive:
598 1 Patrick McGarry
https://www.dell.com/learn/us/en/04/shared-content~data-sheets~en/documents~dell-red-hat-cloud-solutions.pdf 
599 1 Patrick McGarry
http://www.fujitsu.com/global/products/computing/storage/eternus-cd/
600 1 Patrick McGarry
http://www8.hp.com/h20195/v2/GetPDF.aspx/4AA5-2799ENW.pdf http://www8.hp.com/h20195/v2/GetPDF.aspx/4AA5-8638ENW.pdf 
601 1 Patrick McGarry
http://www.supermicro.com/solutions/storage_ceph.cfm 
602 1 Patrick McGarry
https://www.thomas-krenn.com/en/products/storage-systems/suse-enterprise-storage.html
603 1 Patrick McGarry
http://www.qct.io/Solution/Software-Defined-Infrastructure/Storage-Virtualization/QCT-and-Red-Hat-Ceph-Storage-p365c225c226c230 
604 1 Patrick McGarry
605 1 Patrick McGarry
Notices:
606 1 Patrick McGarry
Copyright © 2016 Intel Corporation. All rights reserved
607 1 Patrick McGarry
Intel, the Intel logo, Intel Atom, Intel Core, and Intel Xeon are trademarks of Intel Corporation in the U.S. and/or other countries. *Other names and brands may be claimed as the property of others.
608 1 Patrick McGarry
Results have been estimated based on internal Intel analysis and are provided for informational purposes only. Any difference in system hardware or software design or configuration may affect actual performance.
609 1 Patrick McGarry
Intel® Hyper-Threading Technology available on select Intel® Core™ processors. Requires an Intel® HT Technology-enabled system. Consult your PC manufacturer. Performance will vary depending on the specific hardware and software used. For more information including details on which processors support HT Technology, visit http://www.intel.com/info/hyperthreading.
610 1 Patrick McGarry
Any software source code reprinted in this document is furnished under a software license and may only be used or copied in accordance with the terms of that license.
611 1 Patrick McGarry
INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL PRODUCTS. NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL'S TERMS AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE OF INTEL PRODUCTS INCLUDING LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT. 
612 1 Patrick McGarry
613 1 Patrick McGarry
A "Mission Critical Application" is any application in which failure of the Intel Product could result, directly or indirectly, in personal injury or death. SHOULD YOU PURCHASE OR USE INTEL'S PRODUCTS FOR ANY SUCH MISSION CRITICAL APPLICATION, YOU SHALL INDEMNIFY AND HOLD INTEL AND ITS SUBSIDIARIES, SUBCONTRACTORS AND AFFILIATES, AND THE DIRECTORS, OFFICERS, AND EMPLOYEES OF EACH, HARMLESS AGAINST ALL CLAIMS COSTS, DAMAGES, AND EXPENSES AND REASONABLE ATTORNEYS' FEES ARISING OUT OF, DIRECTLY OR INDIRECTLY, ANY CLAIM OF PRODUCT LIABILITY, PERSONAL INJURY, OR DEATH ARISING IN ANY WAY OUT OF SUCH MISSION CRITICAL APPLICATION, WHETHER OR NOT INTEL OR ITS SUBCONTRACTOR WAS NEGLIGENT IN THE DESIGN, MANUFACTURE, OR WARNING OF THE INTEL PRODUCT OR ANY OF ITS PARTS. 
614 1 Patrick McGarry
615 1 Patrick McGarry
Intel may make changes to specifications and product descriptions at any time, without notice. Designers must not rely on the absence or characteristics of any features or instructions marked "reserved" or "undefined". Intel reserves these for future definition and shall have no responsibility whatsoever for conflicts or incompatibilities arising from future changes to them. The information here is subject to change without notice. Do not finalize a design with this information. 
616 1 Patrick McGarry
617 1 Patrick McGarry
The products described in this document may contain design defects or errors known as errata which may cause the product to deviate from published specifications. Current characterized errata are available on request. 
618 1 Patrick McGarry
619 1 Patrick McGarry
Contact your local Intel sales office or your distributor to obtain the latest specifications and before placing your product order.