태그 : 위치로그 : 방명록 : 관리자 : 글쓰기  
 
Study, Study, Study!
공부하자!
  카테고리
전체 (102)
cache (51)
interc... (2)
schedu... (13)
OS Issues (5)
power (7)
simula... (1)
Memory (1)
Archit... (17)
ILP an... (3)
applic... (1)
  공지사항
  달력
«   2012/01   »
1 2 3 4 5 6 7
8 9 10 11 12 13 14
15 16 17 18 19 20 21
22 23 24 25 26 27 28
29 30 31        
  태그
  최근에 올라 온글
: Reducing Cache Misses Usin...
: dynamic hardware-assisted ...
: Better than the Two : Exce...
: SOS: A Software-Oriented D...
: Fairness via Source Thrott...
  최근에 달린 댓글
: 잘 지내긴 -_- 읽기만 ...
: 은지야. 애-_- 키우다 ...
: 이번 주 연구실 세미나...
: 앗 감사합니다. 블로그 ...
: PACT 08 에서 DIP의 단...
: simple is the best라는...
: MapReduce를 보면서, 특...
: 오늘은 내가 졸려서 내...
: DATE, 2006
: MICRO'2006
  최근에 받은 트랙백
  링크
  글 보관함
Total : 57934
Today : 84
Yesterday : 175
Get RSS Page XML RSS2.0
Powered By TatterTools
Skin By Foradun.com
Edit By Badung.net
 
2011/08/01 23:38 2011/08/01 23:38
 Reducing Cache Misses Using HW and SW Page Placement, ICS'1999 - cache/partitioning : 2011/08/01 23:38
single core의 page coloring

code와 data page는 conflict가 많다, 에서 출발.

먼저 SW로는,
compiler level에서 code와 data page의 conflict matrix를 그려서 OS에 hint로 주고,
OS에서는 이를 보고 conflict가 적게 나도록 page coloring을 수행한다.

그리고 HW로,
recoloring을 빨리 하기 위해 TLB에 remap field를 두어서 recolor.

성능 향상 20% 이상. .
관련글0 : 댓글0






2011/08/01 21:03 2011/08/01 21:03
 dynamic hardware-assisted software-controlled page placement to manage capacity allocation and sharing within large caches, HPCA'2009 - 분류없음 : 2011/08/01 21:03
page migration overhead를 줄이기 위해 adress bit 중 안쓰는 bit 이용.
mapping을 한 단계 더 두는 방식으로 조절.
page re-coloring overhead를 줄이고 싶을 때 채택할 수 있는 방법 중 하나라고 생각.

관련글0 : 댓글0






2011/06/28 20:19 2011/06/28 20:19
 Better than the Two : Exceeding Private and Shared Caches via 2-Dim Page Coloring, HPCA2007 - cache/shared vs private : 2011/06/28 20:19

기존에 MICRO'2006 에 tile구조에서 위치를 가깝게 하기 위해
SW-level에서 control한 것에 이어서 paper를 확장한 것.

tile구조에서, 가까운 곳에 place하게 하면서,
가까운 곳 중에서도, conflict가 가장 작을 곳에 place.
conflict는 program-level에서 각 page별 conflict를
matrix형태로 표현하고 그를 page assign할 때 적용.

이렇게 하면 hybrid를 coloring을 통해 수행 가능.

관련글0 : 댓글0






2011/06/24 22:57 2011/06/24 22:57
 SOS: A Software-Oriented Distributed Shared Cache Management Approach for CMP, PACT2009 - cache/shared vs private : 2011/06/24 22:57
CMP의 Last Level Cache의 Private/Shared hybrid 방식.
shared는 평균 latency가 길고, private은 한 core가 쓸수 있는 capacity가 적다.
장단점을 hybrid한 paper들이 많았는데, 이 paper는 그를 OS-level에서 해결한 work.

multi-threaded application이 target이고
compile time에 off-line profiling 을 통해서 각 data를 어떻게 place할지 결정한다.
access pattern별로 placement를 다르게 하기 위해서
even partition, scattered, dominant owner, small-entity, shared로 classification을 한다.
classification을 위해서는 K-means clustering algorithm을 사용.
관련글0 : 댓글0






2011/06/24 19:08 2011/06/24 19:08
 Fairness via Source Throttling, ASLPOS'2010 - scheduling : 2011/06/24 19:08
Onur Mutlu group. 지난번 학교에서 이 분 세미나 들었었는데
intensive하게 연구하는 분 중 한분. 젊으시다. 그리고 잘생겼다. +_+

fairness를 얻기 위한 방법은 여러 가지가 있는데
cache partitioning을 할 수도 있고, frequency scailing을 할 수도 있고.
근데 여기서는 특이하게 MSHR를 조절한다.

한 쪽의 performance가 더 높거나 해서 fair하지 않다고 판단될 때
L2 cache입장에서 그 안에서 partitioning을 하는 방법도 있겠지만.
source에 해당하는 MSHR에서 message를 보내는 비율을 줄이면
L2 cache에서 자연스럽게 cache를 차지하는 비율도 줄어들 것이다.

참 괜찮은 생각인 것 같은 idea였다.
관련글0 : 댓글0







↑top