Component reliability questions:
1. DRAM
1a. DRAM soft errors --> between 1-10 FITs / Gbit [Charles Slayman, SUN, IRPS 2008]
(1 FIT = 1 error per billion device hours)
64 GBytes / DRAM server for 10K servers we obtain 512*10K *(1 to 10) FITs
total --> Mean time to errors roughly 20 to 200 hours (assumes all
flips are important).
ECC necessary.
DRAM vendors are seeing peripheral logic upsets inside DRAM chips
and that error rate stays roughly constant on a per bit basis --> they will be
significant contributors --> even if I take their contribution to be only 10%
still significant --> traditional ECC doesn't do anything about these errors --
need to mix address with data --> need a quick study to see if error detection
will be enough --> needs support for recovery / retry.
More questions: S/w techniques on commodity DRAM or special DRAM?
(interesting techniques available)
Scrubbing (relevant for the next discussion on hard errors).
1b. DRAM hard errors: There is increasing concern about hard errors in DRAMs AND
alignment of hard and soft errors. This concern was raised by IBM, and
a 2008 IBM R&D by TJ Dell.
Why is Chipkill necessary? -- IBM data: (with simple parity:
7 fails per 100 servers with 32MB mem. with parity over 3 yrs.,
9 fails per 100 servers with 1 GB mem. with SEC over 3 years)
Chipkill alone may not be enough in that case
and has to be backed up by either clever ECC or scrubbing (not sure scrubbing
will be enough).
According to the Dell paper, there is a significant
probability of DRAM chips producing hard errors (not a single
hard error but entire chip or most part of it). If not repaired right away,
there can be significant contribution to mem. failure rate (I THINK
the numbers below are for a 32 GB mem. subsystem but we need to
doublecheck the math)
Time to repair (months) Memory failure rate adder (FITs)
--------------------------------------------------------------------------------
1 102
6 608
12 1,207
Solutions: scrubbing enough? -- it depends
Redundant mem. modules? -- expensive?
If we know some DRAM chip is BAD, we can use erasure
codes (but it may not be that straightforward -- we need to think).
Theremay be several other opportunities -- e.g., if we replicate the
data for a variety of reasons -- redundancy to prevent large outages / data
loss, performance -- then we should be able to play other tricks -->
But error DETECTION should be crucial --> (simple parity won't suffice
for alignment reasons unless we take preventive actions by redistributing
data instead of waiting for repair) --> the numbers in the above table
become less important at that point.
There is a recent paper at SIGMETRICS 09
that discusses these hard errors as well --
it seems they are also seeing increasing incidence of DRAM hard errors.
(DRAM Errors in the Wild: A Large-Scale Field Study, Bianca
Schroeder, Eduardo Pinheiro, Wolf-Dietrich Weber, SIGMETRICS, 2009 (to
appear:
Here are some excerpts from the abstract:
They analyzed measurements of memory errors in a large fleet
of commodity servers over a period of 2.5 years. The collected data
covers multiple vendors, DRAM capacities and technologies, and
comprises many millions of DIMM days. They observed
DRAM error rates that are orders of magnitude higher than
previously reported, with 25,000 to 70,000 errors per billion device
hours per Mbit and more than 8% of DIMMs affected by errors per year.
They have strong evidence that memory errors are dominated by hard
errors, rather than soft errors).
2. Storage servers:
Assuming 10K storage servers --> will be interesting both from system
building / modeling / measurement purposes as well as to see what
protection techniques will be useful (or if any will be necessary).
Even if we assume that each server contributes 100 FITs for SOFT ERRORS (transients)
(the number is not off -->
and most possibly this number takes into account derating at the node level
(probability that a flip-flop error doesn't cause a system error -- this is generally taken into account when vendors quote error rates).
Overall: 10^6 FITs: MTTF of 1,000 hrs.: assumes full usage and
all applications highly critical --> which won't be the case in real life probably.
Needs characterization of workload types, their criticality, and a full system
derating (note: data replication doesn't solve this problem --> server errors
different)
Plus, one has to worry about Hard errors: increasing: aging, etc..
(even if you have 20-50 FITs per chip for hard errors --> can add up).
Opportunities:
Criticality of applications will be key -- e.g., social networking sites
may not care much for silent errors --> other data center apps (e.g., banking
or commercial) will.
Protection techniques -- we will discuss a few (software only?
hardware-assisted techniques -- probably not an option for our COTS implementation;
However, we should investigate those).
Software techniques --> scrubbing alone
won't work --> error detection (any application-level properties for end-to-end
or time redundacy based? --> performance impact, energy impact?
Also, brings up the question of:
checkpointing, recovery support?
Are we going to have transactions semantics? -- Can that help?
Can we classify transactions as "critical" vs. "non-critical" and
rely on selective protection?
For hard errors --> very thorough on-line self-test / on-line self-diagnostics
can PREDICT (i.e., EARLY detection even before errors appear)
failures: generally requires hardware support: may not
be available on our COTS parts --> may be interesting experiment opportunities.
Brings up questions of self-repair / self-healing --> what if hard errors are DETECTED
(instead of predicted) by periodic self-diagnostics --> implications on recovery?
Also, relevant: interactions with power management, power overhead
of error checking.
3. Application servers
--> much relaxed reliability requirements? (social
networking -- they often use commodity app. servers and focus on storage
servers?).
Need to address -- criticality of app: selective protection -->
can be carried over through the entire system (our previous discussion
on storage servers).
Another issue: who is to be blamed for incorrect (criticality wise) results
due to app. server errors.
4. Networking substrate:
I missed the last part of the discussion
on whether we need custom networking substrate to meet our
latency objectives or not -- if we do need custom substrates,
that can provide opportunities in doing stuff from reliability standpoint.
If not, one must analyze error rates of the switches as well.
Some vendors thought that their error rates would be really
low because of CRC checks providing end-to-end checks only to
find later that CRC wasn't done right to protect errors in switches
themselves. We can project numbers similar to our earlier discussions
for these as well.
System reliability questions:
I haven't considered questions such as power outages,
disasters, geographical diversity, etc. in detail -- reason: for such
causes, what is important is data replication -- detection is less of an
issue -- issue is recovery) -- we can discuss that as part of this discussion.
However, this part is closely related to what we need to do for DRAM
(hard and soft) errors.
I looked up a few recent papers on causes of system failure rates.
Here is some data:
1. Understanding Failures in Petascale Computersby Bianca Schroeder Garth A. Gibson
Data set during 1995-2005 at LANL:
22 HPC systems (4,750 machines, 24,101 processors).
22 clusters (18 SMP-based clusters, 2 to 4 processors per
node: Total 4,672 nodes, 15,101 processors; remaining 4: NUMA
boxes with 128 to 256 processors each: total 78 nodes, 9,000 processors).
An entry for any failure that occurred the time period
that resulted in an application interruption or a node outage.
(Note: it seems they are not doing concurrent checking so I'd assume
silent data corruption isn't included here).
System failure causes covered: software failures, hardware
failures, failures due to operator error, network failures, and failures due to environmental problems (e.g. power outages).
Here is some interesting data:
Cluster node outages: > 50% -- hardware failures; ~ 20% software causes;
15% -- unknown; rest: network, human, environment causes (doublechecked
by looking at the fraction of repair time attributed to these causes).
Number of failures per year per system varies between 20 to 1,100 per
system --> on an average: 0.2 to 0.7 per system per year per processor.
(Note: as the paper suggests, their systems mainly implement checkpointing; don't do much about error detection).
(There has been some reported data that Blue Gene systems have
significantly lower failure rates --> need to know details).
Cray XT3/XT4: 10,880 CPUs, Failures per month
per TF: 0.1 to 1
IBM Power 5/6: 10,240 CPUs, 1.3
Clusters AMD x86: 8,000 CPUs, 2.6 to 8
BlueGene L/P: 131,720 CPUs, 0.01 to 0.03
(this is taken from:
H. Simon. Petascale computing in the U.S.. Slides from presentation at the ACTS workshop.
http://acts.nersc.gov/events/Workshop2006/slides/Simon.pdf, June 2006).
Next paper:
Tahoori, Kaeli & others: they looked at storage servers
(so far as I know it's EMC).
Normalized failure rates:
Hardware-related 1.91 (System A) 2.27 (System B1) 7.25 (System B2)
2.19 (Total)
Power-related 0.18 (A) 0.19 (B1) 0.5 (B2) 0.19 (Total)
Software-related 2.41 (A) 4.44 (B1) 18.12 (B2) 3.48 (Total)
SEU-related 1.0 (A) 1.0 (B1) 1.0 (B2) 1.0 (Total)
Of course there are several papers (including Jim Gray 1985
paper which discusses operator errors being significant, Microsoft
XP paper discussing importance of third party drivers, etc.).