Oracle SQL Parsing Flow Diagram(SQL 解析流程图)

       在MOS 上有一篇介绍SQL 解析流程图的文档,参考[ID32895.1]。 截取主要部分内容如下:

               Thisarticle show the parsing mechanism as a flow diagram. Its main purpose is toshow the difference between a 'soft' and a 'hard' parse. It is intended to givea feel of how parsing operates to make explanation of parsing activity easier.

 

 

 

 

注意事项:

1. A cursor is an address on the client that points tothe memory location of a SQL statement on the server. Multiple-client cursors may point at the same address on the server.

 

2. Remember that 'Client' and'Server' sides may reside on the same machine - in which caseClient/Server is a logical distinction.

 

3. If a cursor is open, then thestatement will be in the sql_area, so no parsing is necessary. This iswhy locks may remain when a client is terminated abnormally (such as a PCClient being turned off without closing open cursors).

 

4. SESSION_CACHED_CURSORS is the initialisation parameterthat specifies how many cursors to hold open for a particular session. The opencursor request will still be sent to the server but it will not be executedonce a matching cursor is found in the session cursor cache.

 

5. HOLD_CURSOR is an precompiler parameter that specifiesthat an individual cursor should be held open.

 

6. Both the soft and hard parseregister as a parse in tkprof. Hashing the current statement updates theparse count.

 

 

7. Soft parse avoids many of thesteps taken during the parse phase for a particular statement. Initialsyntactic and semantic checks are made and then the statement is hashed andcompared with hashed statements in the SQL area. If amatch is found, then existing information is used and relatively expensivesteps (such as query optimization etc.) areavoided.

 

8. The 10053 event is only invoked during a hard parse.

 

Oracle 10053 事件

http://www.cndba.cn/Dave/article/1570

 

 

 

 

-------------------------------------------------------------------------------------------------------

QQ:492913789

Email:ahdba@qq.com

Blog:  http://www.cndba.cn/dave

Weibo:   http://weibo.com/tianlesoftware

Twitter:  http://twitter.com/tianlesoftware

Facebook: http://www.facebook.com/tianlesoftware

Linkedin: http://cn.linkedin.com/in/tianlesoftware

DBA1 群:62697716(满);   DBA2 群:62697977(满)  DBA3 群:62697850(满) 

DBA 超级群:63306533(满);  DBA4 群: 83829929  DBA5群: 142216823   

DBA6 群:158654907  聊天 群:40132017   聊天2群:69087192

--加群需要在备注说明Oracle表空间和数据文件的关系,否则拒绝申请

内容概要:本文详细比较了GPU、TPU专用AI芯片在大模型推理优化方面的性能、成本及适用场景。GPU以其强大的并行计算能力和高带宽显存,适用于多种类型的神经网络模型和计算任务,尤其适合快速原型开发和边缘计算设备。TPU专为机器学习设计,擅长处理大规模矩阵运算密集型任务,如Transformer模型的推理,具有高吞吐量和低延迟特性,适用于自然语言处理和大规模数据中心的推理任务。专用AI芯片通过高度定制化架构,针对特定神经网络模型进行优化,如卷积神经网络(CNN),在处理特定任务时表现出色,同时具备低功耗和高能效比的优势,适用于边缘计算设备。文章还介绍了各自的优化工具和框架,如CUDA、TensorRT、TPU编译器等,并从硬件成本、运营成本和开发成本三个角度进行了成本对比。 适合人群:从事人工智能、深度学习领域的研究人员和技术人员,尤其是对大模型推理优化感兴趣的读者。 使用场景及目标:①帮助读者理解GPU、TPU和专用AI芯片在大模型推理中的优缺点;②为选择适合的硬件平台提供参考依据,以实现最优的推理性能和成本效益;③介绍各种优化工具和框架,帮助开发者高效部署和优化模型。 其他说明:本文不仅涵盖了硬件架构特性,还深入探讨了优化技术和应用场景,旨在为读者提供全面的技术参考。在选择硬件平台时,需综合考虑具体任务需求、预算限制及开发资源等因素。
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值