K.L.E.

Society’s
COLLEGE OF ENGINEERING & TECHNOLOGY
Udyambag – Belgaum – 590008
(An ISO 9001-2000 Certified Institution)

Seminar Report On
A Phase Change Memory as a Secure Main Memory

Submitted by,
Rutuja Ladge
[USN: 2KL07CS058]

Department of Computer Science & Engineering

Academic Year 2010-11

Department of Computer Science Engineering

CERTIFICATE

This is to certify that “Name” bearing USN:2KL07CS058 has satisfactorily given the

seminar on “A Phase Change Memory as a Secure Main Memory” during academic year
2010-2011 in partial fulfillment of the requirements for the award of degree of Bachelor of
Engineering in Computer Science and Engineering by Visvesvaraya Technological
University, Belgaum.

Co-ordinator H.O.D
 Abstract

Phase change memory (PCM) technology appears as more scalable than DRAM technology. As PCM
exhibits access time slightly longer but in the same range as DRAMs, several recent studies have proposed to
use PCMs for designing main memory systems. Unfortunately PCM technology suffers from limited write
endurance; typically each memory cell can be only being written a large but still limited number of times (107
to 109 writes are reported for current technology). Till now, research proposals have essentially focused their
attention on designing memory systems that will survive to the average behavior of conventional applications.
However PCM memory systems should be designed to survive worst-case applications, i.e., malicious attacks
targeting the physical destruction of the memory through overwriting a limited number of memory cells. In this
paper, a design is proposed of a secure PCM-based main memory that would by construction survive to
overwrite attacks.
Contents
Chapter 1
1. Introduction
1.2 Motivation
1.3 Literature survey and introduction

Chapter 2
2.1 A secure PCM based main memory ……………………………….11
2.1 Security principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . …….11
2.1.1 Invisible PA-to-PCMA translation is required . . . . . . . . . . . . . . . .11
2.1.2 PA-to-PCMA translations must dynamically . . . . . . . . . . . . . . . . .11
2.2 Principles of a practical secure PCM-based main memory . . . . . . . .11
2.2.1 PA-to-PCMA translation . . . . . . . . . . . . . . . . . . . . . . . . . . . ……..11
2.2.2 PA-to-PCMA region address translation . . . . . . . . . . . . . . . . . . …11
2.2.3 PA-to-PCMA region displacement translation . . . . . . . . . . . . . . . ..11
2.2.4 Dynamically changing PA-to-PCMA translation . . . . . . . . . . . . . . 12
2.2.5 How to modify PA-to-PCMA translation . . . . . . . . . . . . . . . . . .

5 PCM memory controller ………………………………………………..13

Chapter 3
3.1 Applications
Chapter 4
4 Future trends in PCM technology(Comparison and evaluation)
4.1 Economic feasibility of PCM main memory . . . . . . . . . . . . . . . . . . . . . 15
4.2 Page mode is compatible with security . . . . . . . . . . . . . . . . . . . . . . . …15
4.3 Limiting extra write traffic overhead . . . . . . . . . . . . . . . . . . . . . . . . ….15
5 Conclusion
1 Introduction
Phase Change Memory (PCM) technology appears as a promising technology for designing main memory in
future computer systems. PCM presents advantages over DRAMs in terms of static energy consumption as well
as integration scalability for future technologies generations; for instance, anticipates durance, i.e., a PCM
memory cell can only support a limited number of writes and exceeding this limit might impair its correct
functioning. The reported write endurances for PCM memory vary between 107 and 109 writes on a single cell.
Such a limited endurance has been recognized as issue for the design of PCM-based main memory systems.
Several propositions have been made to allow a PCM main memory to survive the anticipated lifetime of a
computer system ,i.e., 10 to 20 years, in the context of general applications. At the exception of these studies
completely ignore the security breach that the limited write endurance of PCM components would create in a
main memory. PCM components for main memory would create a main memory through a very simple
program overwriting the same memory cells again and again. The potential attack is particularly simple to
mount. It can be run by any user without any execution privilege. It is seen that their Region Based Start Gap
scheme would survive a few months to a naive overwrite attack consisting in constantly overwriting the same
physical memory address. However, the Region Based Start Gap (RBSG) scheme considered in would not
survive more than a few days to a slightly more complex attack based on the birthday paradox. More over the
RBSG scheme a RBSG scheme supporting page mode would even be less endurant to an overwrite attack. This
paper proposes the design of a secure main PCM memory. In order to prevent a malicious user to overwrite
some memory cells, the physical memory address (PA) manipulated by the computer system is not the same as
the PCM memory address (PCMA) . PCMA is made invisible from the rest of the computer system. The PCM
memory controller is in charge of the PA-to-PCMA translation. Hiding PCMA alone does not prevent a
malicious user to blindly overwrite some PCM memory blocks. Therefore in the secure PCM-based main
memory, PA-to-PCMA translation is continuously modified through a random process. This prevents a
malicious user to overwrite some PCM memory words, it also uniformizes the write pressure on the overall
memory for every possible type of workloads. For implementing the PA-to-PCMA translation, the PCM
memory controller implements a translation table and needs an efficient random number generator. As an
example, for write endurance in the 32M range, our study shows that associating a single translation table entry
with a 4K memory blocks region should be sufficient. Provided one extra write per 8 program generated writes,
our scheme would resist an overwrite attack for 62 % of the expected total memory lifetime. However,
endurance to overwrite attacks is obtained at the cost of some performance decrease on applications limited by
the main memory bandwidth since one extra block read and one extra memory block write is generated every
eight memory block writes. The security also limits the number of possible program-generated writes on the
memory to 8/9 th of the total number of possible writes on the memory.
3 PCM attributes
This new class of non-volatile memory brings together the best attributes of NOR, NAND and RAM.
1. Bit-alterable:
Like RAM or EEpROM, PCM is bit alterable. Flash technology requires a separate erase step in order to change
information. Information stored in bit-alterable memory can be switched from a one to zero or zero to a one
without a separate erase step.
2. Non-volatile:
Like NOR ash and NAND ash, PCM is nonvolatile. RAM, of course, requires a constant power supply, such as
a battery backup system, to retain information. DRAM technologies also suffer from susceptibility to so-called
"soft errors" or random bit corruption caused by alpha particles or cosmic radiation. Early testing results
conducted by Intel on multimegabit PCM arrays for long term data retention show excellent results.
3. Read speed:
Like RAM and NOR-type ash, the technology features fast random access times. This enables the execution of
code directly from the memory, without an intermediate copy to RAM. The read latency of PCM is comparable
to single bit per cell NOR ash, while the read bandwidth can match DRAM. In contrast, NAND ash suffers from
long random access times on the order of 10s of microseconds that prevent direct code execution.
4. Write/erase speed:
PCM is capable of achieving write speeds like NAND, but with lower latency and with no separate erase step
required. NOR ash features moderate write speeds but long erase times. As with RAM, no separate erase step is
required with PCM, but the write speed (bandwidth and latency) does not match the capability of RAM today.
The capability of PCM is expected, however, to improve with each process erase speed PCM is capable of
achieving write speeds like NAND, but with lower latency and with no separate erase step required. NOR ash
features moderate write speeds but long erase times. As with RAM, no separate erase step is required with
PCM, but the write speed (bandwidth and latency) does not match the capability of RAM today. The capability
of PCM is expected, however, however, to improve with each process generation as the PCM cell area
decreases. Scaling is the fifth area where PCM will offer a difference. Both NOR and NAND rely on memory
structures which are difficult to shrink at small lithos. This is due to gate thickness remaining constant and the
need for operation voltage of more than 10V while the operation of CMOS logic has been scaled to 1V or even
less. This scaling effect is often referred to as Moore's Law, where memory densities double with each smaller
generation. With PCM, as the memory cell shrinks, the volume of GST material shrinks as well, providing a
truly scalable solution. PCM employs a reversible phase change phenomenon to store information through a
resistance change in different phases of a material. An advance in memory technology and pioneering work
conducted by Numonyx has moved the technology to the forefront of the memory industry R & D activity.
PCM offers a combination of some of the best attributes of NOR ash, NAND ash, EEpROM and RAM in a
single memory device. These capabilities uniquely combined with the potential for lower memory subsystem
costs could potentially create new applications and memory architectures in a wide range of systems.
4 A secure PCM based main memory
4.1 Security principles
4.1.1 Invisible PA-to-PCMA translation is required
If a malicious attacker knows the PA-to-PCMA translation then for a given PCM memory block B, he/she is
able to _gure out the address of the physical memory block that is mapped on B. If the PA-to-PCMA translation
is made invisible from the outside of the PCM memory then the attacker can not retrieve the address of the
physical memory block mapped on a given memory block.
4.1.2 PA-to-PCMA translations must dynamically
Our analysis of the RBSG scheme has shown that, in order to resist a birthday paradox attack, the PA to PCMA
translation of any physical block B has to be modi_ed with a frequency largely higher than one time every
Wmax possible writes on B. The PA-to-PCMA translation changes should be completely unpredictable from
the outside of the PCM memory; in particular there should be no restrictions on the new translation.
4.2 Principles of a practical secure PCM-based main memory
4.2.1 PA-to-PCMA translation
In the secure PCM-based main memory the PA-to-PCMA translation is performed by the PCM memory
controller through the use of a translation table. For a physical memory block B, the address of the
corresponding PCM block is computed from an entry read in the translation table and the address B. The PA to
PCMA translation must perform a one-to-one address translation from the physical address space to PCM
address space. The simplest mapping would be to associate a translation table entry with each physical memory
block and ensuring that the translation is a one-to-one block mapping. Such one-to-one block mapping appears
as unpractical so associate a single translation table entry with a region of R contiguous memory blocks; for
instance if 4K contiguous memory blocks are mapped by a single entry, 64K entries are su_cient to map 16
GBytes. Such a single translation for a large region was already proposed in the context of wear leveling for
conventional applications
4.2.2 PA-to-PCMA region address translation
Initializing at boot time the translation table T with a one-to-one region mapping is unpractical. Instead of such
an initialization, assume that at initialization time, the translation table T is initialized with only zeros, but that
some computation is performed at runtime in addition to the read of the translation table. If memory regions are
numbered from 0 to N 􀀀 1, the translation is performed as follows: region B in physical memory is mapped
onto region (T(B):addressxorBxorRinit), where Rinit is a random number generated at initialization time. Note
that, at initialization time,T(B) is null and that the PA-to-PCMA region address translation is a one-to-one
mapping
4.2.3 PA-to-PCMA region displacement translation
The use of a single entry to map a complete region of the physical memory could lead to a possible overwrite
attack on trying to write a specific block in all the regions, for instance the first block. In order to avoid such an
attack, the displacement in the region is also translated. Physical memory block X in region B is mapped onto
block (T(B):dispxorXxorDinit) in region (T(B):addressxorBxorRinit). As Rinit;Dinit is a random number
generated at initialization time.
4.2.4 Dynamically changing PA-to-PCMA translation
In order to avoid blind overwrite attacks, the PA-to- PCMA translation must be continuously modified. More
precisely, only writes represent an issue. Therefore, the PA-to-PCMA translation modification is triggered
randomly and only on memory writes. This random triggering is particularly important: As an example, if the
PA-to-PCMA translation change occurs periodically on writes, for instance every 10 writes, an attacker could
repeat the sequence of nine consecutive writes on physical block B, one write on physical block C. Physical
block C moves in the PCM memory, but block B remains on the same PCM memory location that can be easily
overwritten.
4.2.5 How to modify PA-to-PCMA translation
Modifying the PA-to-PCMA translation for a physical region B is implementing through swapping the
translations for two physical regions. This guarantees that the PA-to-PCMA translation remains a one-to-one
mapping. The region swapping induces the modification of two entries in the translation table. A random
physical regionB0 is chosen and the PA-to-PCMA translations ofBandB0 are exchanged, i.e.,
T(B):addr := old(T(B0):addr) xor B0 xor B
and
T(B0):addr := old(T(B):addr) xor B0 xor B
4.2.5.1 At the same time, displacement translations inside blocks B and B' are also modified,
T(B):disp := old(T(B)):disp xor RAND
and
T(B0):disp := old(T(B0)):disp xor RAND
where RAND is randomly selected. The two memory regions in the PCM memory have to read and swapped
accordingly. Randomly swapping memory regions has been already proposed in the context of wear leveling for
ash memories. Frequency of PA-to-PCMA translation modifications. The cost of a PA-to-PCMA translation
modification is proportional to the size R of a region in the memory. The two swapped regions have to be read
and rewritten, i.e., a PA-to- PCMA translation modification induces 2R memory block reads and 2R memory
block writes. Therefore, the frequency of the address translation modification should be chosen in order to
maintain the total overhead to a reasonable level. This study arbitrarily estimate that inducing in average one
extra write on the PCM memory per 8 effective writes IEEE Computer Architecture Letters Phase Change
Memory (PCM) is a term used to describe a class of non-volatile memory devices that employ a reversible
phase change in materials to store information. Matter can exist in various phases such as solid, liquid, gas,
condensate and plasma. PCM exploits differences in the electrical resistivity of a material in different phases.
This paper describes the basic technology and capabilities of PCM would be acceptable. That is, in average one
out of 16R physical memory block writes can trigger a PA-to-PCMA address translation modification.
Therefore on receiving a write on a physical memory block, the modification of its PA-to-PCMA translation is
randomly triggered with probability 116R.
5.PCM memory controller
The design of a secure PCM-based main memory leads to several constraints inside the memory controller.
5.1 Write endurance and region size
The principles above lead to the design of a PCM-based main memory on which an overwrite attack would only
be able to consecutively write the same memory block in average 16R times before the physical block is moved
in another PCM memory block. In practice, a write attack could succeed in significantly reducing the lifetime of
the memory was done. If 16R is not small with respect to the write endurance of the cells. Simulations of an
overwrite attack on a 16 GBytes PCM memory i.e., 226 256-byte blocks. Regions of respectively 64K and 4K
memory blocks were considered. If the memory features a write endurance of Wmax =2E writes, then the
theoretical write endurance of a uniformly accessed PCM memory is 226+E. With a write endurance of only 8
Megawrites (223) per cell, using 64K memory blocks per region is not an option: some memory blocks would
be destroyed by a brute force overwrite attack in less than a billion (230) writes. With 4K memory blocks
regions, the PCM memory would be able to support an attack consisting of up to 38 % of the theoretical 249
writes. If the write endurance is 32 Mega writes per cell (225) then these respective ratios become 7.4 % for
64K memory blocks regions, and 62 % for 4K memory blocks region for a theoretical maximum of 88.88 %
since in average one extra block write is triggered for 8 physical memory writes. If the write endurance is 256
Mega writes per cell (230) then these ratios become 52 % and 79% for 64K and 4K memory block regions
respectively. Therefore, if the technology is able to ensure write endurance in the hundreds of millions range
then even very large regions could be considered for PA-to-PCM translations.
5.2 Memory controller constraints
The secure PCM main memory would need to integrate extra hardware in the memory controller or implement
the secure PA-to-PCMA translation. Memory storage volume The storage volume of the PCM memory
controller is a major issue. The main component is the translation table that features an entry per memory
region. For a 16 GB memory, the use of regions of 4K 256-bytes blocks would lead to 64K entries, each entry
featuring the address of region (14 bits)and a displacement in the region (12 bits) i.e. a total of 26 bits. The total
storage cost of the translation table would be 52Kbytes. If a 64K blocks region was used then the size of the
PCM translation table would only be 3.25 Kbytes.
5.3 Swapping memory regions logic
The memory controller has to handle the important function of swapping two memory regions on a PA-to-
PCMA translation change. This induces a large number of memory reads and writes. An atomic swap of the two
memory regions would stop the normal read and write accesses by the computer system. This would be
unacceptable. Therefore the memory controller must feature logic to interleave blocks swapping with the
normal flow of reads and writes from the computer system. The logic must be able to handle the case where a
normal write is overwriting a block belonging to one of the memory regions being swapped. Moreover this
normal flow of writes may randomly trigger new region swaps; the memory controller should be able to buffer
these swaps. The priority on writes must be dynamically adapted in order to maintain a limited number of
regions waiting for swaps, for example at most 8 swaps. As an example, The policy of randomly splitting the
write priority to 1/4th for region swapping and 3=4th for normal write was tested. Flow when less than 4 region
swaps are waiting and one half for region swapping and one half for normal write flow when 4 or more region
swaps. On an experiment on 240 writes and assuming a continuous saturated write flow from the computer
system, there was never more than 8 waiting region swaps.
5.4 Extra PA-to-PCMA translation latency
The extra access time to main memory due to PA to- PCMA translation is essentially due to the read of the PA-
to-PCMA translation table. This table will be implemented as a SRAM table in the memory controller. For a 16
GB memory using 4K 256-byte blocks region, the read access time of a 52Kbytes SRAM memory would in the
range of 2-4 processor cycles and would be marginal compared with the overall main memory access time.
5.5 The random number generator
The secure PCM-based main memory will be able to resist to an overwrite attack if no one is able to follow or
reconstruct the PA-to-PCMA translation process. Our proposal heavily relies on a random number generator.
The security of our proposal also depends of the security of this random number generator. One can remark that
the output of the random number generation used in our memory controller cannot be directly observed from
the outside of the PCM memory. Therefore different possible schemes could be implemented ranging from a
true hardware random generator to a simpler algorithmic pseudo-random number generator personalized with a
huge key at manufacturing time.
5.6 Write Endurance for conventional applications
The security of our proposed secure PCM main memory is ensured by
1. the invisibility of the access on the PCM IEEE Computer Architecture Letters memory from outside the
memory, and

2. the random distribution of the accesses through the PA-to-PCMA address translation change: under an
overwrite attack, a given physical block will change PA-to-PCMA address translation after a number of writes
that is comparatively small with the write endurance of the PCM memory cell. For a conventional application,
for a given physical memory region, the frequency of the PA-to-PCMA address translation changes measured in
writes on the region is the same as under an overwrite attack, but the writes are distributed over the whole
region. Therefore, for each block, the number of writes is very small compared with the endurance of the PCM
memory cell. The write endurance of the overall memory system is therefore very close to its maximum
theoretical endurance
6 Future trends in PCM technology
6.1 Economic feasibility of PCM main memory
If within a few years, the write endurance per cell on PCM components reach 256 M writes then it would
become feasible to build 16 GBytes (or larger) memory using comparatively very small PA-to-PCMA
translation table (for instance, using 64K blocks memory regions): as mentioned above, the secure PCM
memory would be able to survive to an overwrite attack at a full 4 GBytes/s write bandwidth for 52
6.2 Page mode is compatible with security
If PCM memories are used as main memory then a page mode would be interesting as on current DRAM to
limit the access latency and increase bandwidth when the memory read requests exhibit high spatial locality.
Our PA-to-PCMA translation scheme is compatible with such page mode since regions are large enough to
accommodate large page, even split across several PCM components.
6.3 Limiting extra write traffic overhead
In this study, the overwrite traffic associated with PA-to-PCMA translation modification could be as large as
1extra PCM block write per 8 physical memory block writes. This overhead can be reduced by decreasing the
probability of triggering a PA-to-PCMA translation. This would reduce the total endurance of the system to the
overwrite attack, but may still remain acceptable if the cell write endurance is large. For instance, for a
16GBytes memory, if the extra PCM write traffic is limited to 1 extra PCM block write per 32 physical memory
writes i.e., a 3.1 % write overhead, the cell endurance is 256 Megawrites and 64K blocks memory regions are
used then the secure PCM memory still survives a full 4 GBs/s bandwidth overwrite attach memory block
writes. This overhead can be reduced by decreasing the probability of triggering a PA-to-PCMA translation.
This would reduce the total endurance of the system to the overwrite attack, but may still remain acceptable if
the cell write endurance is large. For instance, for a 16GBytes memory, if the extra PCM write tra_c is limited
to 1 extra PCM block write per 32 physical memory writes i.e., a 3.1% write overhead, the cell endurance is 256
Mega writes and 64K blocks memory regions are used then the secure PCM memory still survives a full 4
GBs/s bandwidth overwrite attack for 19 % of its 32 years expected lifetime, i.e., about 6 years.
3 Applications
1. PCM in embedded systems:
A common use of memory in an embedded system is for code storage. Systems requiring a relatively small
amount of memory, less than approximately 2Gb, are architected such that code is executed directly from the
NOR ash . This memory is often white paper Phase Change Memory: A new memory to enable new memory
usage models used as storage memory for an embedded _le system. DRAM is often used in these types of
systems as a scratchpad memory. In these types of systems, PCM can be used as a code execution memory.
With its bit-alterable feature, PCM is able to displace some or all of the DRAM required in the system. In
embedded system applications , PCM can reduce the density requirements for DRAM while fulfillling the
density requirement of the NAND ash. At the same time, the presence of PCM in this type of system simplifies
and improves the performance of _le systems stored in the PCM due to the bit-alterability and low latency
features.
2. PCM in wireless systems:
It is common to find nearly independent subsystems for baseband and application processing in wireless
systems. At the highest level, these can be considered as independent embedded systems. Generally speaking,
both subsystems have the need for a resident execution memory and for storage of small data structures. In
many cases, the applications subsystem is also expected to store and perform operations on larger multimedia
content. are slower but on the same order of magnitude as the latencies of DRAM, albeit on smaller page sizes,
PCM can serve as an outstanding code execution memory and outstanding read-mostly memory for all but the
most frequently manipulated data structures. The bit-alterability of PCM eliminates the need for block erase,
which reduces the DRAM requirements even further, resulting in a lower cost memory subsystem. PCM
promises a scalable memory subsystem solution that provides the best overall cost while meeting the increasing
performance demands of high-end, multimedia wireless devices.
3. PCM IN SOLID STATE DEVICES
Managing NAND ash in solid state storage (SSD) subsystems is a challenge due to the block-alterable nature of
the NAND technology. It is also challenging to handle increasing levels of error management required when the
memory is heavily program/erase cycled or frequently read. PCM can be used in SSD systems to store
frequently accessed pages and to store those elements which are more easily managed when manipulated in
place. Examples of these types of elements include: parity bits for data stored in NAND, bad block tables, and
block and page mapping tables. In this scenario, a small amount of PCM could be used to enhance the
manageability of NAND. By minimizing the stress on the NAND memory, higher density MLC NAND is
enabled, thus leveraging the capability of PCM to lower the cost of the NAND ash in the subsystem. This
caching with PCM will improve the performance and reliability of the subsystem. that multiple erase cycles are
likely required to free space to store the new data being written to the device. This increases the number of
cycles on the device and further accelerates the time until the maximum endurance limits are reached. The bit-
alterable nature of PCM solves the issue of increased write cycles when the device is full. Higher endurance of
PCM addresses the needs of these systems when heavy use is expected.PCM in computing platforms As a
volatile memory, DRAM consumes power to simply maintain the contents of the memory.
4. PCM as non-volatile memory:
PCM banks can be turned off when they are not in use to provide reduced power in idle states. More
importantly, turning off the banks decouples the relationship between density and power consumption. This
results in a PCM subsystem density envelope that is not limited by the power envelope constraints of that
system. In addition to non-volatility, PCM offers endurance and write latencies that are compelling for this type
of application. This is a key advantage over read-mostly solutions that have been attempted until now.
5 Conclusion
If the promises of the PCM technology are ful_lled then it might become economically feasible to build a main
memory from PCM memory component in the next few years. Such a PCM-based main memory will
particularly be attractive due to its very low static energy consumption. However to consider such a memory for
an industry product , the PCM based memory would have to be able to resist to software overwrite attacks
targeting its physical destruction. In this, a first secure PCM based main memory that will resist to overwrite
attacks . By hiding the effective PCM memory address from the rest of the computer system and continuously
and randomly moving the physical memory blocks in PCM memory, overwrite attacks are made impossible.
The proposed PA-to-PCMA translation scheme uniformizes and randomizes the write flow on PCM memory
for malicious overwrite attacks as well as conventional non malicious applications. Our scheme requires some
hardware overhead in the memory controller (essentially a PA-to-PCMA translation table, the memory region
swapping logic and a random number generator). But our scheme allows the overall PCM main memory to
resist to an overwrite attack for a very significant fraction of its expected lifetime, e.g. 62 % for 4Kblocks
regions and 32M write endurance per cell. The scheme presented in this paper induces a significant extra write
traffic (1 extra write per 8 effective writes). This consumes some useful memory write bandwidth and may
reduce performance on memory intensive applications. A future study will address this issue for conventional
(i.e., non-malicious attack) applications.